The Workshop Programme

The Workshop Programme
The Workshop Programme
9:30- 9:35 Introduction
9:35-10:00 Penny Labropoulou, Harris Papageorgiou, Byron Georgantopoulos, Dimitra
Tsagogeorga, Iason Demiros and Vassilios Antonopoulos:
"Integrating Language Technology in a web-enabled Cultural Heritage system"
10:00-10:30 Oto Vale, Arnaldo Candido Junior, Marcelo Muniz, Clarissa Bengtson, Lívia
Cucatto, Gladis Almeida, Abner Batista, Maria Cristina Parreira da Silva, Maria
Tereza Biderman and Sandra Aluísio:
"Building a large dictionary of abbreviations for named entity recognition in
Portuguese historical corpora"
11:00-11:30 Eiríkur Rögnvaldsson and Sigrún Helgadóttir:
"Morphological tagging of Old Norse texts and its use in studying syntactic
variation and change"
11:30-12:00 Lars Borin and Markus Forsberg:
"Something Old, Something New: A Computational Morphological Description of
Old Swedish"
12:00-12:30 Dag Trygve Truslew Haug and Marius Larsen Jøhndal:
"Creating a Parallel Treebank of the Old Indo-European Bible Translations"
12:30-12:55 Elena Grishina:
"Non-Standard Russian in Russian National Corpus (RNC)"
12:55-13:20 Dana Dannélls:
"Generating tailored texts for museum exhibits"
14:30-15:00 David Bamman and Gregory Crane:
"The Logic and Discovery of Textual Allusion"
15:00-15:30 Dimitrios Vogiatzis, Galanis Dimitrios, Vangelis Karkaletsis and Ion
Androutsopoulos:
"A Conversant Robotic Guide to Art Collections"
15:30-16:00 René Witte, Thomas Gitzinger, Thomas Kappler and Ralf Krestel:
"A Semantic Wiki Approach to Cultural Heritage Data Management"
16:30-17:15 Christoph Ringlstetter:
"Error Correction in OCR-ed Data"- Invited Talk
17:15-17:30 Closing
i
Workshop Organisers
Caroline Sporleder (Co-Chair), Saarland University, Germany
Kiril Ribarov (Co-Chair), Charles University, Czech Republic
Antal van den Bosch, Tilburg University, The Netherlands
Milena P. Dobreva, HATII, University of Glasgow, Scotland
Matthew James Driscoll, Kobenhavns Universitet, Denmark
Claire Grover, University of Edinburgh, Scotland
Piroska Lendvai, Tilburg University, The Netherlands
Anke Luedeling, Humboldt-Universitat, Germany
Marco Passarotti, Universita Cattolica del Sacro Cuore, Italy
Workshop Programme Committee
Ion Androutsopoulos, Athens University of Economics and Business, Greece
Timothy Baldwin, University of Melbourne, Australia
David Bamman, Perseus, USA
David Birnbaum, University of Pittsburgh, USA
Antal van den Bosch, Tilburg University, The Netherlands
Andrea Bozzi, ILC-CNR, Pisa, Italy
Kate Byrne, University of Edinburgh, Scotland
Paul Clough, Sheffield University, UK
Greg Crane, Perseus, USA
Milena P. Dobreva, HATII, University of Glasgow, Scotland
Mick O'Donnell, Universidad Autonoma de Madrid, Spain
Matthew James Driscoll, Kobenhavns Universitet, Denmark
Franciska de Jong, University of Twente, The Netherlands
Claire Grover, University of Edinburgh, Scotland
Ben Hachey, University of Edinburgh, Scotland
Djoerd Hiemstra, University of Twente, The Netherlands
Dolores Iorizzo, Imperial College London, UK
Christer Johansson, University of Bergen, Norway
Piroska Lendvai, Tilburg University, The Netherlands
Anke Luedeling, Humboldt-Universitat, Germany
Roland Meyer, University of Regensburg, Germany
Maria Milosavljevic, University of Edinburgh, Scotland
Marie-Francine Moens, Katholieke Universiteit Leuven, Belgium
Marco Passarotti, Universita Cattolica del Sacro Cuore, Italy
Martin Reynaert, Tilburg University, The Netherlands
Kiril Ribarov, Charles University, Czech Republic
Maarten de Rijke, University of Amsterdam, The Netherlands
Peter Robinson, ITSEE, UK
Maria Simi, University of Pisa, Italy
Caroline Sporleder, Saarland University, Germany
ii
Table of Contents
David Bamman and Gregory Crane:
"The Logic and Discovery of Textual Allusion"
1
Lars Borin and Markus Forsberg:
"Something Old, Something New: A Computational Morphological Description of
Old Swedish"
9
Dana Dannélls:
"Generating tailored texts for museum exhibits"
17
Elena Grishina:
"Non-Standard Russian in Russian National Corpus (RNC)"
21
Dag Trygve Truslew Haug and Marius Larsen Jøhndal:
"Creating a Parallel Treebank of the Old Indo-European Bible Translations"
27
Penny Labropoulou, Harris Papageorgiou, Byron Georgantopoulos, Dimitra
Tsagogeorga, Iason Demiros and Vassilios Antonopoulos:
"Integrating Language Technology in a web-enabled Cultural Heritage system"
35
Eiríkur Rögnvaldsson and Sigrún Helgadóttir:
"Morphological tagging of Old Norse texts and its use in studying syntactic
variation and change"
40
Oto Vale, Arnaldo Candido Junior, Marcelo Muniz, Clarissa Bengtson, Lívia
Cucatto, Gladis Almeida, Abner Batista, Maria Cristina Parreira da Silva, Maria
Tereza Biderman and Sandra Aluísio:
"Building a large dictionary of abbreviations for named entity recognition in
Portuguese historical corpora"
47
Dimitrios Vogiatzis, Galanis Dimitrios, Vangelis Karkaletsis and Ion
Androutsopoulos:
"A Conversant Robotic Guide to Art Collections"
55
René Witte, Thomas Gitzinger, Thomas Kappler and Ralf Krestel:
"A Semantic Wiki Approach to Cultural Heritage Data Management"
61
iii
Author Index
Almeida, Gladis
Aluísio, Sandra
Androutsopoulos, Ion
Antonopoulos, Vassilios
Bamman, David
Batista, Abner
Bengtson, Clarissa
Biderman, Maria Tereza
Borin, Lars
Candido Junior, Arnaldo
Crane, Gregory
Cucatto, Lívia
Dannélls, Dana
Demiros, Iason
Dimitrios, Galanis
Forsberg, Markus
Georgantopoulos, Byron
Gitzinger, Thomas
Grishina, Elena
Haug, Dag Trygve Truslew
Helgadóttir, Sigrún
Jøhndal, Marius Larsen
Kappler, Thomas
Karkaletsis, Vangelis
Krestel, Ralf
Labropoulou, Penny
Muniz, Marcelo
Papageorgiou, Harris
Rögnvaldsson, Eiríkur
Silva, Maria Cristina Parreira da
Tsagogeorga, Dimitra
Vale, Oto
Vogiatzis, Dimitrios
Witte, René
47
47
55
35
1
47
47
47
9
47
1
47
17
35
55
9
35
61
21
27
40
27
61
55
61
35
47
35
40
47
35
47
55
61
iv
The Logic and Discovery of Textual Allusion
Gregory Crane
The Perseus Project
Tufts University
Medford, MA
[email protected]
David Bamman
The Perseus Project
Tufts University
Medford, MA
[email protected]
Abstract
known to the reader, we are using it here in the specific context of an imitative textual allusion – a passage in one text that refers to a passage in another.
When Willy Loman calls each of his sons an “Adonis” in Death of a Salesman, there is no doubt that
this is an allusion to a Classical myth, but it does not
point to a definable referent in the record of written humanity (as King’s allusion refers specifically
to the first six words of the Gettysburg Address).
The discovery of these allusions is a crucial process for the analysis of texts. As others have pointed
out,1 allusions have two main functions: to express
similarity between two passages, so that the latter
can be interpreted in light of the former; and to simultaneously express their dissimilarity as well, in
that the tradition they recall is revised.2 Allusions of
this specific variety are perhaps most widely known
as a trope of modernist authors such as Eliot and
Joyce, but they are common in the Classical world
as well – most strongly in the Greek poetry of the
Hellenistic era, in the Roman poetry of the republic
and early empire and in New Testament texts (which
allude to prophecies recorded in the Old Testament).
Given the long history of Latin literature, we must
also keep in the mind a text’s Nachleben – how it
has been received and appropriated by the generations that follow it.3
Uncovering allusions of this sort has long been
the task of textual commentators, but we present
We describe here a method for discovering
imitative textual allusions in a large collection of Classical Latin poetry. In translating
the logic of literary allusion into computational terms, we include not only traditional
IR variables such as token similarity and ngrams, but also incorporate a comparison of
syntactic structure as well. This provides a
more robust search method for Classical languages since it accomodates their relatively
free word order and rich inflection, and has
the potential to improve fuzzy string searching in other languages as well.
1
Introduction
Five score years ago, a great American, in whose symbolic shadow we stand
today, signed the Emancipation Proclamation ...
Thus begins Martin Luther King Jr.’s “I Have a
Dream” speech of 1963. While the actual text of the
Gettysburg Address is not directly quoted here, it is
elicited by means of an allusion: King’s audience
would immediately have recognized the parallels between his first four words and the “Four score and
seven years ago” that began Lincoln’s own speech.
By opening with this phrase, King is aligning Lincoln’s invocation of human equality with “the greatest demonstration for freedom in the history of our
nation” for which he was then speaking.
While the term “allusion” is commonly applied
to any reference to a person, place, or thing already
1
For an overview of the function and interpretive significance of allusions, see Thomas (1986).
2
Cf. Bloom (1973).
3
Cicero, for example, was widely admired by Renaissance
humanists after Petrarch and provided a model for textual imitation. Cf. Kristeller (1979).
1
a method here to automatically discover them in
texts. Our approach has many similarities with research on text reuse (Clough et al., 2002), paraphrase and duplicate detection (Dolan et al., 2004),
and locating textual reference (Takeda et al., 2003;
Lee, 2007), but while these methods generally focus
on string comparison and document structure, we include variables for considering the abstract structure
of a sentence as well, as represented by its syntax.
This enables a more robust search method since it is
not restricted by word order or inflection. Our test
corpus is a collection of Latin poetry, but the methods we describe are language independent.
2
month, a week, a natural day, That Faustus may
repent and save his soul! O lente, lente, currite noctis equi! (Act V, Scene 2)
And again, four centuries later, Vladimir
Nabokov appropriates it for Lolita as his protagonist
is chased along a highway.
(3) We were many times weaker than his splendid,
lacquered machine, so that I did not even attempt to outspeed him. O lente currite noctis
equi! O softly run, nightmares! (Nabokov 219)
Following Irwin (2001), we can distinguish an allusion from a mere quotation in the level of context required to understand it. A quotation is selfcontained; an allusion calls forth the original context in which it’s found. Direct allusions like these
are easier to find than their adapted counterparts (it
is essentially a simple string search) but they reside
on the same continuum as the others.
Types of Textual Allusion
While others have categorized textual allusion into
a number of types dependent on their function (e.g.,
Thomas (1986) distinguishes between “casual reference,” “single reference,” “self-reference,” etc.),
we are concerned only with a practical distinction
in terms of the ease of locating them: an allusion is
either direct (equivalent to a quotation) or indirect.
2.1
2.2
Indirect reference
Most of what we would consider allusions involve
some transformation of the referent text. An example of this can be found in the first line of the first
poem of Ovid’s Amores, an imitation (and revision)
of the first line of Vergil’s Aeneid.
Direct reference
The most explicit and unambiguous type of allusion
is direct reference in the form of a verbatim quotation. We see this form of allusion most often in the
long afterlife of a text, as for instance in the reception of this line from Ovid’s Amores.
(4) Arma gravi numero violentaque bella parabam
/ Edere (Am. 1.1-2)5
(5) Arma virumque cano (Aen. 1.1)6
Vergil’s Aeneid is an epic poem focussed on the
figure of Aeneas (an ancestor of the Romans), written in dactylic hexameter, the same “heavy” meter as
Homer’s epics the Iliad and Odyssey. Ovid, in contrast, is a love poet, and elicits Vergil’s famous opening to motivate his genre (the line continues with Cupid stealing one of the line’s metrical feet, leaving it
an elegaic couplet, a common meter of Roman love
poetry).
This type of common allusion clearly presents
much more difficulty in being found: any variety of
simple string search (either exact or fuzzy) will not
be successful, since only two word forms – arma
(“arms”) and the enclitic -que (“and”) – are common
to both strings.
(1) At si, quem mavis, Cephalum conplexa teneres
/ Clamares: lente currite, noctis equi! (Am.
1.13)4
While Ovid’s line comes from the mouth of the
mythic Aurora (dawn) pleading with her chariot to
pull her more slowly across the sky to give her more
time with her lover before returning to her husband,
Christopher Marlowe sixteen centuries later appropriates it for Faust, who voices it in the final minutes
before midnight in a plea to prolong his life.
(2) Stand still, you ever-moving spheres of heaven,
That time may cease, and midnight never come:
Fair Nature’s eye, rise, rise again and make Perpetual day; or let this hour be but A year, a
5
“I was planning to write about arms and violent wars in a
heavy meter.”
6
“I sing of arms and the man.”
4
“But if you held Cephalus in your arms, whom you prefer,
you would shout ‘run slowly, horses of the night!”’
2
3
mar.7 In both of these trees, the two direct objects of
the verbs are headed by the coordinator que via the
syntactic relation OBJ CO, while the coordinator is
headed by the verb via the relation COORD. While
the words themselves vary, the structure is the same.
The Logic of Allusion
Clearly we need to add new methods for establishing
similarity between two lines beyond simple string
matches. This begs the question, however, of how
it is we know (as humans) that one passage in a
text is an allusion to another. The ultimate criterion
of course involves higher-order reason (an allusion
must make interpretive sense) but we can identify a
number of explicit surface variables that give notice
to the presence of an allusion in the first place.
(edere)
que
COORD
arma
OBJ_CO
Identical words. A quotation is an allusion where
the edit distance between two strings is effectively
0: i.e., all word forms in one span of text are identical with those in another. In sentences 4 and 5,
only arma and que are the same, but they nevertheless provide a necessary anchor for establishing a
link between the two passages. While arma in both
examples here in is the same grammatical case (accusative), many times an alternation occurs as well
(e.g., transforming a word from the accusative to the
nominative case). We can therefore define “identical” to mean both token identity (arma = arma) and
root form (lemma) identity (ego = me).
(bella)
OBJ_CO
Figure 1: Dependency tree of arma -que bella edere
(“to write about arms and wars”).
(cano)
que
COORD
arma
OBJ_CO
(virum)
OBJ_CO
Figure 2: Dependency tree of arma virumque cano
(“I sing of arms and the man”).
Word order. Syntax in projective languages like
English is strongly tied to word order (an adjective,
for example, generally modifies the noun that immediately follows it), but for non-configurational languages like Latin and Greek, word order is much
more free, especially in the genre of poetry in which
allusion is so common. For this reason we treat syntax as a separate variable (see below) and isolate
word order as its own phenomenon. For our example above, word order is another cue to the presence
of an allusion since both lines begin with the same
word, arma.
Metrical and phonetic similarity The first lines
of both of these poems are both written in dactylic
hexameter, but the similarity between the two is
much closer than that, since the first seven syllables
of both lines are metrically identical – two dactyls
followed by a stressed syllable and caesura. Additionally, the final long syllable before the caesura is
the same in both sentences (“o”), eliciting a further
phonetic similarity as well.
Syntactic similarity. When considering syntax
we begin to see the strongest parallels between the
two passages. In both sentences, arma is involved in
coordination as a direct object of a verb. While the
head verbs differ (edere vs. cano) as does the other
object involved in coordination (bella vs. virum), the
two structures are syntactically identical.
Figures 1 and 2 present a syntactic tree of each
sentence under the formalism of dependency gram-
Semantic similarity We can also note that on a
semantic level, both of these passages are “about”
similar things, at least in this first line (before the
arrival of Cupid in Ovid) – in both lines, the author
is communicating (via writing or singing) about war
(violenta bella) and the instruments of war (arma).
(6) Ārmă grăvı̄ nŭmĕrō
(7) Ārmă vı̆rūmqŭe cănō
...
...
7
This is the structural representation of syntax as found in
the Latin Dependency Treebank (Bamman and Crane, 2007)
and the Prague Dependency Treebank of Czech (Hajič, 1998).
3
cos(~s, ~t) = pPn
pP
Here si is the tf/idf score for the term i in the
source sentence s and ti is the tf/idf score for that
same term in the target comparsion sentence t. We
measure each tf/idf by the following formula.
N
(1 + log(tfi,j )) log df
i
Here tfi = the count of term i in sentence j, N =
the total number of sentences in the collection, and
dfi = the number of sentences in that collection that
contain the term i.
The closer this cosine is to 1, the more similar two
sentences are. We will use this general framework to
inform all of the following variables: the difference
between them will be in what exactly constitutes a
“term.”
(8) The Moons resplendent Globe (PL 4.723)
(9) Lucentemque globum lunae (Aen. 6.725)
4.1
These five categories represent broad classes of
similarity, but of course we must expect others on
an ad hoc basis as well – in sentences 4 and 5
from above, we have the additional similarity that
both passages come from the privileged first lines of
both poems, suggesting a larger structural similarity.
While these variables do not illuminate the interpretive significance of an allusion (we can leave that
contentious task to critics), they do provide a means
by which to discover them in the first place.
4
Pn
st
i=i i i
n
2
s
t2
i=i i
i=i i
With semantic similarity we can also group another very important variable – cross-language semantic information in the form of translation equivalents. This is extremely important given the reception of these texts across cultures and distant
eras. Classical Roman poets themselves are especially fond of borrowing from Homer and Hellenistic poets, but we see the same phenomenon in English as well – one only need to look at Milton’s use
of the Aeneid in Paradise Lost to see the level of appropriation, which in its simplest form approaches
exact translations of fixed phrases, such as sentences
8 and 9 below, and in its more complex form also involves the host of other variables outlined above.
Identical words
Given Latin’s rich inflection, we will define two
variables for establishing identity between words,
token similarity and lemma similarity.
Token similarity. Here we define term to be the
overt (i.e., inflected) form of the word. This measure
reflects a typical search engine query in that it compares two documents (here, sentences) based on how
closely their words match each other. More common
words between the two documents leads to a greater
level of similarity.
Discovering Allusions
Lemma similarity. Here we define term to be the
uninflected lemma from which the token is derived.
In this variable, omnia vincit amor (“love conquers
all”) is identical to omnia vincuntur amore (“all
things are conquered by love”) since the lemmas underlying both are omnis1 vinco1 amor1. A measure
for lemma similarity addresses the fact that many allusions are not simple quotations – the words that
constitute the reference are not bound to their original case as they were used in the target text, but are
often given a different grammatical role in the allusion.
Our task in automatically discovering allusions is to
transform the variables listed above into ones that
are computationally tractable. We need to be able
to define the precise degree to which two passages
are similar in order to quantitatively compare which
pairs of passages are more similar to each other than
others.
Information retrieval has produced a number of
metrics for judging the similarity of two documents.
The most widely used of these generally assign a
relevance score based on some variation of tf/idf
weighting: two documents are similar if they both
contain words that occur less frequently in the collection of texts at large. The more uncommon words
they share, the greater their similarity.
To establish the similarity between two sentences,
we can use the cosine measure as a means of judging
their vector similarity.
4.2
Word order
We can measure the explicit order of words (as distinct from their abstract syntax) with the use of ngrams – specifically bigrams and trigrams, which
measure how frequently two or three words appear
4
eludet
PRED
in linear order. Using the beginning and end of sentences as distinct words of their own (in order to
measure when a word begins or ends a line), the
phrase omnia vincit amor has 4 bigrams ([start] omnia, omnia vincit, vincit amor, and amor [end]) and
three trigrams: ([start] omnia vincit, omnia vincit
amor, and vincit amor [end]).
This will let us capture, for instance, that arma
virumque cano is similar to arma gravi numero in
that both begin with the bigram [start] arma. We
can again account for Latin’s rich inflection with the
use of lemma bigrams and trigrams in addition to
tokens. This results in four total word order variables: token bigram, token trigram, lemma bigram
and lemma trigram.
4.3
furor
SBJ
iste
ATR
nos
OBJ
tuus
ATR
Figure 3: Dependency tree of furor iste tuus nos
eludet (“that madness of yours will play with us”),
Cicero, In Catilinam 1.1.
themselves. How we represent those words constitutes the first dimension:
• part of speech: adj:noun:verb
Syntax
• token: iste:furor:eludet
The two variables outlined so far form the backbone
of information retrieval applications. By considering syntax, we can get beyond simple string resemblance metrics and begin to consider similarities in
abstract structure as well.
With syntactic relations, we can specify the true
syntactic distance between two phrases (as distinct from simple word order). Several measures
of syntactic distance have recently been proposed:
Spruit (2006) presents a method for classifying dialects based on previously human-curated variables
(e.g., the presence of personal vs. reflexive pronouns etc.); Nerbonne and Wiersma (2006) approximate syntactic distance using part of speech
trigrams, which works well for classifying different language groups (adults vs. child) in English
(a language with strict word order); and Sanders
(2007) measures distance using Sampson’s (2000)
leaf-ancestor paths, in which each word in a sentence is identified as its path from itself to the
top of the syntactic tree (e.g., in a phrase structure
grammar: “The”-Det-NP-S/“dog”-N-NP-S/“barks”V-VP-S). Given Latin’s non-projectivity, we have
adopted this third measure and augmented it along
three dimensions to make it suitable for a dependency grammar.
Figure 3 presents a syntactic tree annotated under
a dependency-based grammar. Since dependency
grammars do not have intermediate phrase structures
such as NP or VP, we take our basic syntactic structure to be a child-parent relationship between words
• lemma: iste1:furor1:eludo1
The second dimension is the length of the path.
While Sanders’ metric identifies each word by its
full path from itself to the top of the tree, we can use
a number of intermediary paths to assert similarity
as well. Since a full path from the word itself to
the top of the tree is very unlikely to be repeated
across sentences, we approximate it by considering
only paths of lengths 2 and 3 (bigrams and trigrams):
a path of length 2 would for instance be comprised of
“adj:noun”/“iste:furor”/etc. while a path of length 3
would be comprised of “adj:noun:verb” (as above).
The third dimension is the presence or absence of the syntactic label. Dependency grammars differ from phrase structure grammars by providing an explicit relation between words (where
phrase structure grammars often imply it by position – a subject, for example, is an NP that ccommands a VP). Using the syntactic labels specified in the Latin Dependency Treebank (Bamman
and Crane, 2007), a labeled path would be comprised of “iste:ATR:furor:SBJ:eludet” for token trigrams, while an unlabeled path would leave this information out (as above).
These three dimensions provide 12 distinct syntactic variables for each word in a sentence, ranging from least explicit (unlabeled part of speech bigrams [“adj:noun”]) to most (labeled token trigrams
[“iste:ATR:furor:SBJ:eludet”]). The most explicit
5
Author
Ovid
Vergil
Horace
Catullus
Propertius
variables will have the lowest inverse document frequencies and will therefore be the most informative
for judging similarity if present, while the least explicit variables will still provide a back-off means to
provide some similarity in the event of a more explicit non-match.
4.4
Metrical/phonetic similarity and semantic
similarity
Sentences
10,459
6,553
2,345
903
366
20,626
Table 1: Composition of the test corpus by author.
While we do not implement metrical/phonetic or semantic similarity measures in what follows, we can
address the means by which we could do so in the
future.
We can measure metrical and phonetic similarity in a manner similar to the term frequencies used
in the variables above, by comparing the meter of
two passages (this of course requires metrically annotated texts). Meter in this case can be seen as a
language with two letters, long( ) and short( ), and
˘
¯
we can judge the similarity between two meters as a
simple string comparison of that representation.
We can judge the semantic similarity between
two words using either monolingual clustering techniques such as latent semantic analysis (which notes,
for example, that an apple is semantically close to an
orange since both appear often with words such as
eat and tree) (Deerwester et al., 1990), or by crosslanguage translation equivalents (such as those induced in the course of parallel text alignment (Och
and Ney, 2003)), which notes the frequency with
which a word in one language (such as oratio in
Latin) is translated by different terms (e.g., speech
vs. prayer).
5
Words
141,091
97,495
35,136
14,793
4,867
293,382
al.’s dependency parser (McDonald et al., 2005) on
the manually curated data in the Latin Dependency
Treebank and used it to parse all of the texts in our
collection.8
After finding the most similar sentences for each
of the 20,626 sentences in our collection, we filtered the results to require a lower limit for sentence
length in order to find meaningful pairs (short sentences such as Quid est? can be found across many
authors and are not allusions even though they match
exactly) and to avoid sentence pairs that are both
found in the same immediate context (e.g., Catullus’ poem 61, where a chorus of the same 7 words is
exactly repeated 11 times throughout the poem).9
The results are encouraging: while a detailed
quantitative evaluation must await the creation of a
test corpus of canonical allusions, we can at least
now provide a list of the closest matches for all sentences in our collection. For any given sentence, further research will of course be necessary to discern
whether it represents a real allusion, but the highest
scoring pairs in our experiment tend to be strong examples. Sentences 10 and 11, for instance, present
one such pair from Ovid and Vergil with a similarity
score of .173.
Evaluation
We evaluated the first three variable classes above
(word identity, word order and syntax) on a collection of 14 texts from 5 Latin poets – Catullus
(Carmina), Ovid (Metamorphoses, Amores, Epistulae, Medicamina Faciei Femineae, Ars Amatoria,
Remedia Amoris), Vergil (Aeneid, Eclogues, Georgics), Propertius (Elegies I), and Horace (Carmina,
Satyrarum libri, De Arte Poetica liber).
While the word identity and word order variables
can be calculated on unstructured text, we need syntactically parsed data in order to measure syntactic
similarity. To create this, we trained McDonald et
(10) Innumeras urbes atque aurea tecta videbis,
/ Quaeque suos dicas templa decere deos (Ov.
Ep. 16)10
(11) Iam subeunt Triviae lucos atque aurea tecta
(Verg., Aen. 6.13)11
8
In a tenfold test on the treebank data itself, we measured
the parser’s unlabeled accuracy to be 64.99% and its labeled
accuracy to be 54.34% (Bamman and Crane, 2008).
9
o Hymen Hymenaee io, o Hymen Hymenaee.
10
“You will see innumerable cities and golden roofs, and tempes that you would say are fitting to their gods.”
11
“Already they enter Trivia’s groves and golden roofs.”
6
Variable
nullae:puellae:ATR
nullae:puellae
nulli/illum
p:SBJ EXD OBJ CO:u:COORD:v
,/nullae
nullus1:puella1
nullus1:puella1:ATR
nullae
...
nulli
puellae
illum
a:n:ATR:v:SBJ
Sentences 12 and 13 likewise present a pair from
Ovid and Catullus with a score of .141.
(12) nulli illum iuvenes, nullae tetigere puellae
(Ov., Met. 3.353)12
(13) idem cum tenui carptus defloruit ungui / nulli
illum pueri, nullae optavere puellae (Cat.,
Carm. 62)13
The strongest matches, however, came within authors, who often sample their own work in other
contexts. This occurs most often by far in Vergil,
where the re-appropriation involves exactly repeating complete sentences (9 instances), exactly repeating substantial sentence fragments (23 instances),14
and more significant modifications.
Additionally, since our weights are based on preset variables, the process by which we come to the
most similar match is transparent. Table 2 presents
the term weights for several of the highest and lowest variables at play in establishing the similarity between sentences 12 and 13 above.
This table presents the clear importance of using
syntax as a method for establishing the similarity between sentences – three of top four variables that
have linked these two sentences to each other involve syntax (e.g., nullae depends on puellae in both
sentences as an attribute).15
Our search for loci similes to our original allusion from above – Ovid’s Arma gravi numero violentaque bella parabam – illustrates well the importance of bringing a variety of information to the
search. The closest sentences to Ovid’s original line
all bear some similarity to it on both a lexical and
syntactic level (as sentences 1 and 2 demonstrate below). Our target sentence of Vergil (Arma virumque
cano ...), however, only shows up in 11th place on
the list.
tf/idf
9.24
9.24
9.24
9.24
8.84
8.55
8.32
8.55
...
6.30
5.55
5.34
1.67
Table 2: Sample of variable contribution. Components separated by a colon represent syntactic relations; those with slashes are n-grams.
1. Arma procul currusque virum miratur inanes (.059) (Aen. 6.651)16
2.
Quid tibi de turba narrem
numeroque virorum (.042) (Ov., Ep.
16.183)17
11. Arma virumque cano, Troiae
qui primus ab oris Italiam, fato profugus,
Laviniaque venit litora, multum ille et terris iactatus et alto vi superum saevae memorem Iunonis ob iram (.025) (Aen. 1.1)18
This is understandable given the variables we
have implemented – the first three sentences do indeed bear a closer similarity to the original without
being diluted by extra words (since our cosine value
normalizes for sentence length). We hope in the future to be able to include other important variables
(such as metrical similarity) as well.
12
“No youths, no girls touched him.”
“This same one withered when plucked by a slender nail;
no boys, no girls hope for it.”
14
Here “substantial” means at least seven consecutive identical words.
15
Note that the labeled syntactic bigram nullae:puella:ATR
has the same tf/idf score as the unlabeled nullae:puellae since
all instances of nullae depending on puella in our automatically
parsed corpus do so via the relation ATR.
13
16
“At a distance he marvels at the arms and the shadowy chariots of men.”
17
“What could I tell you of the crowd and the number of
men?”
18
“I sing of arms and the man, who first from the borders of
Troy, exiled by fate, came to the Lavinian shores – much was he
thrown about on land and sea by force of the gods on account
of the mindful anger of cruel Juno.”
7
6
Conclusion
1990. Indexing by latent semantic analysis. Journal of the American Society of Information Science,
41(6):391–407.
Allusion is by nature an oblique art; its very essence
– referring to something that the audience already
knows – gives it the opportunity to be highly economical in its expression. Since even a single word
or structure can refer to another text, we must leverage as many different varieties of information as we
can in order to discover them, from lexical information to syntax and beyond. We have defined five different variable classes that contribute to the surface
realization of allusion, and have implemented a system that includes three of those five. By considering
the abstract structure of sentences, we are able to effectively search Latin without being encumbered by
its flexible word order and rich inflectional morphology, which allows similar sentences to be expressed
in a variety of ways. While we have designed this
method for a collection of Classical texts, we expect
that it can also be used to improve the robustness of
searches in any language.
7
Bill Dolan, Chris Quirk, and Chris Brockett. 2004. Unsupervised construction of large paraphrase corpora:
exploiting massively parallel news sources. In COLING ’04. Association for Computational Linguistics.
Jan Hajič. 1998. Building a syntactically annotated
corpus: The Prague Dependency Treebank. In Eva
Hajičová, editor, Issues of Valency and Meaning.
Studies in Honor of Jarmila Panevová, pages 12–19.
Prague Karolinum, Charles University Press.
William Irwin. 2001. What is an allusion? The Journal
of Aesthetics and Art Criticism, 59:287–297.
Paul Oskar Kristeller. 1979. Renaissance Thought and
Its Sources. Columbia University Press, New York.
John Lee. 2007. A computational model of text reuse
in ancient literary texts. In Proceedings of the ACL,
pages 472–479, Prague, Czech Republic.
Ryan McDonald, Fernando Pereira, Kiril Ribarov, and
Jan Hajič. 2005. Non-projective dependency parsing
using spanning tree algorithms. In Proceedings of the
HLT-EMNLP, pages 523–530.
Acknowledgments
Vladimir Nabokov. 1991. The Annotated Lolita. Edited,
with preface, introduction, and notes by Alfred Appel,
Jr. Vintage Books, New York.
Grants from the Digital Library Initiative Phrase
2 (IIS-9817484), the National Science Foundation
(BCS-0616521) and the Andrew W. Mellon Foundation (#40700635) provided support for this work.
Thanks are due also to Meg Luthin and Molly Miller
for their invaluable research assistance.
John Nerbonne and Wybo Wiersma. 2006. A measure
of aggregate syntactic distance. In Proceedings of the
Workshop on Linguistic Distances, pages 82–90.
Franz Josef Och and Hermann Ney. 2003. A systematic comparison of various statistical alignment models. Computational Linguistics, 29(1):19–51.
References
Geoffrey Sampson. 2000. A proposal for improving the
measurement of parse accuracy. International Journal
of Corpus Linguistics, 5(1):53–68.
David Bamman and Gregory Crane. 2007. The Latin
Dependency Treebank in a cultural heritage digital
library. In Proceedings of the Workshop on Language Technology for Cultural Heritage Data (LaTeCH 2007), pages 33–40, Prague.
Nathan C. Sanders. 2007. Measuring syntactic difference in British English. In Proceedings of the
ACL2007 Student Research Workshop, pages 1–6.
David Bamman and Gregory Crane. 2008. Building a
dynamic lexicon from a digital library. In Proceedings
of the 8th ACM/IEEE-CS Joint Conference on Digital
Libraries (JCDL 2008).
Marco Rene Spruit. 2006. Measuring syntactic variation
in Dutch dialects. Literary and Linguistic Computing,
21:493–506.
Masayuki Takeda, Tomoko Fukuda, Ichiro Nanri,
Mayumi Yamasaki, and Koichi Tamari. 2003. Discovering instances of poetic allusion from anthologies of
classical Japanese poems. Theoretical Computer Science, 292(2):497–524.
Harold Bloom. 1973. The Anxiety of Influence; A Theory
of Poetry. Oxford University Press, New York.
Paul Clough, Robert J. Gaizauskas, Scott S. L. Piao, and
Yorick Wilks. 2002. METER: Measuring text reuse.
In Proceedings of the ACL, pages 152–159.
R. F. Thomas. 1986. Vergil’s Georgics and the art of
reference. Harvard Studies in Classical Philology,
90:171–98.
Scott C. Deerwester, Susan T. Dumais, Thomas K. Landauer, George W. Furnas, and Richard A. Harshman.
8
Something Old, Something New:
A Computational Morphological Description of Old Swedish
Lars Borin, Markus Forsberg
Språkbanken, Department of Swedish Language
University of Gothenburg
Box 200, SE-405 30 Gothenburg, Sweden
[email protected], [email protected]
Abstract
We present a computational morphological description of Old Swedish implemented in Functional Morphology. The objective of the
work is concrete – connecting word forms in real text to entries in electronic dictionaries, for use in an online reading aid for students
learning Old Swedish. The challenge we face is to find an appropriate model of Old Swedish able to deal with the rich linguistic variation
in real texts, so that word forms appearing in real texts can be connected to the idealized citation forms of the dictionaries.
1.
Background
in digital form on the internet. Two sites with extensive
collections of Old Swedish texts are Språkbanken (the
Swedish Language Bank; <http://spraakbanken.gu.se>)
at the University of Gothenburg and Fornsvensk textbank
(the Old Swedish Text Bank; <http://www.nordlund.lu.
se/Fornsvenska/Fsv%20Folder/index.html>) at Lund
University. The standard reference dictionaries of Old
Swedish have also been digitized (by Språkbanken) and
are available for word lookup via a web interface <http://
spraakbanken.gu.se/fsvldb/> and in full-text for research purposes by special agreement.
1.1. Motivation
Languages with a long written tradition have accumulated
over the centuries a rich cultural heritage in the form of
texts from different periods in the history of the language
community. In these texts, we find information on many
aspects of the origins and history of our culture.
Since languages change over time, older texts can be
difficult or impossible to understand without special training. Indeed, the oldest extant texts in many European languages must in fact be translated in order to be accessible
to a modern reader. In Sweden, the training of professionals who can keep this aspect of our cultural heritage alive by
conveying the content of, e.g., the oldest Swedish legal and
religious texts (13th–15th c.) to a modern audience, is the
province of the academic discipline of Swedish Language
and Linguistics at Swedish universities.
Consequently, an important component of all Swedish
Language university curricula in Sweden is the study of the
older stages of the language. For obvious reasons, any form
of any language older than a bit over a century will be accessible only in writing. Hence, our main source of historical Swedish language data is in the form of old texts, and
courses in the history of the Swedish language, in comparative Scandinavian and in Old Swedish, all contain – in timehonored fashion – a module where students are required
to read a certain amount of texts. In equally time-honored
fashion, students have two main tools for working with Old
Swedish texts: dictionaries and grammar books.
There are excellent editions of Old Swedish texts available in book form, as well as good grammatical descriptions (Noreen, 1904; Wessén, 1969; Wessén, 1971; Wessén,
1965; Pettersson, 2005) and reference dictionaries (section 1.4). Both text collections and reference works such
as dictionaries and grammars tend to be out-of-print works,
which does not in any way detract from their usefulness
or their accuracy, but which does present some practical
problems. Students are confined to using reference copies
held in departmental libraries, but often departments see
themselves forced to restrict access because of the excessive wear this causes.
However, historical texts are increasingly available also
1.2.
Toward a Solution
Our project aims to aid online reading of Old Swedish texts
(cf. Nerbonne and Smit (1996)) by providing access to automatic morphological analysis and linking that to available
lexical resources.
For our goals, we need morphological analysis, i.e. an
analysis module which returns, crucially, a lemma as part
of its result, since the lemma is necessary in order to access the online dictionaries. This makes the methodology
of our project different from that of, e.g., Rayson et al.
(2007), where a POS tagger trained on Modern English
was adapted to deal with the Early Modern English of
Shakespeare’s time. It is doubtful whether this methodology would have helped much if the target language had
been Old English (Anglo-Saxon) instead, because of the
much greater linguistic distance between the two varieties,
which are most appropriately seen as two entirely different, but related languages. What we have in our case are
two historical language stages as far apart linguistically as
Modern English and Anglo-Saxon (see section 1.3).
POS tagging is also a different kind of analysis from
that provided by a morphological analyzer. The former will
provide only one analysis of each text token, the most probable one in the given context, whereas the latter will provide all readings licensed by the lexicon and grammar, but
will not take context into account. POS tagging will always provide an analysis, however, whereas a morphological analyzer may fail to do so if a lemma is lacking in
its lexicon or a form is not allowed by its grammar. Thus,
POS tagging and morphological analysis are complemen-
9
tary; POS tagging is often used to select among competing morphological analyses, the combination of the two
methods thus providing both disambiguation and lemmatization. Our goal in the project described here is to provide (undisambiguated) morphological analysis including
lemmatization of text words, in order to link them to the
corresponding dictionary entries. A natural future extension of the work would be a disambiguation module, either
a POS tagger or, e.g., a constraint grammar (Karlsson et
al., 1995), but we believe that the differences between Old
and Modern Swedish are great enough to preclude an easy
adaptation of a Modern Swedish POS tagger or constraint
grammar.
We are partway through the project. The morphological analysis module is completed for all regular paradigms
and some others (section 2), and a small number of lexical entries have been provided with inflectional information
(section 2.3). This is primarily what we report on in this paper. We are now in the process of providing all lexical entries with inflectional information in the form of paradigm
identifiers (section 2). This task is not completely trivial,
since the extant texts do not always allow us to determine
to which inflectional class a particular lexical entry should
belong.
However, the main challenge still remaining is the issue of how to deal in as principled a way as possible with
the considerable linguistic variation present in the texts that
we are working with, and which presents us with a different situation compared to working with modern texts (sections 1.3 and 3).1
ways, even on the same page of a document. Not only is
there variation in the orthography itself, but also geographical variation, because no unified standard variety had been
established at the time when the texts were produced.
The second factor makes itself felt in the number of
variant forms in our inflectional paradigms (section 2).
As for the third factor, during the second half of the
Old Swedish period the language underwent a development
from the Old Norse (or Modern Icelandic, or Old English)
mainly synthetic language type to the present, considerably
more analytical state. In addition (or perhaps compounding
this process), the sound system of Swedish was thoroughly
reorganized.
For instance, in the nouns, the case system changed profoundly during this period, from the old four-case system
(nominative, accusative, dative, genitive, in two numbers)
to the modern system with a basic form and a genitive clitic
which is the same in all declensions (as opposed to the
old system where there were a number of different genitive markers), and where most functions that the older case
forms expressed by themselves have been taken over by a
combination of free grammatical morphemes and a much
more rigid constituent order.
In the texts that interest us, these changes are in full
swing, which manifests itself as variation in inflectional
endings and in the use of case and other inflectional categories and in the distribution of the corresponding forms.
It is not always easy to tease out the contributions of
these different factors to the linguistic motley evinced by
the texts. Without doubt, the diachronic component is important – the texts are after all from a period three centuries
in length – but it is also probable that the lack of standardization simply allows normal synchronic language variation
to “shine through” in the texts, as it were, rather than being
eliminated as is normally the case with modern, normalized
written standard languages.
1.3.
Down the Foggy Ruins of Time:
Orthographic Inconsistency, Linguistic Variation
and Language Change in Old Swedish Texts
The texts are from the so-called Old Swedish period
(1225–1526), conventionally subdivided into Classical Old
Swedish (1225–1374) and Late Old Swedish (1375–1526),
covering a time span of 300 years. The language of the extant Old Swedish texts exhibits considerable variation, for
at least the following reasons:
1.4.
The Reference Dictionaries
The three main reference dictionaries of Old Swedish are:
• Söderwall (1884) (23,000 entries);
1. The orthography was not standardized in the way that
we expect of modern literary languages;
• Söderwall (1953) (21,000 entries); and
2. the language itself was not standardized, in the sense,
e.g., that a deliberate choice would had been made
about which of several competing forms should be
used in writing; and
• Schlyter (1887) (10,000 entries)
The overlap between the three dictionaries is great, so
that we are actually dealing with less than 25,000 different
headwords. On the other hand, compounds – whether written as one word or separately – are not listed as independent headwords, but as secondary entries under the entry of
one of the compound members. Thus, a full morphological
description reflecting the vocabulary of the three dictionaries will contain many more entries, possibly by an order of
magnitude more.
As an example of the kind of information that is available in the dictionaries, we will briefly discuss the entries
for the word fisker (Eng. ‘fish’), as it appears in these dictionaries. The entry fisker in Söderwall (1884) is shown in
Figure 1. From this entry we learn that fisker is a masculine
noun (indicated by “m.”, in the second line of the entry),
3. the Middle Ages was a time of rapid language change
in Swedish, perhaps more so than any subsequent period of similar length.
The first shows itself in a great variation in spelling;
the “same” word can be spelled in a number of different
1
Of course we are quite aware of the fact that spelling variation
is an empirical fact of modern language as well. There would be
no need of spellcheckers otherwise. In the case of Old Swedish,
however, there were no spelling norms, as far as we can tell. At the
most, there were local scribal practices in the monasteries (which
was where most of the text production occurred), different in different places and never codified, to our knowledge.
10
Figure 1: The entry fisker (Eng. ’fish’) in Söderwall’s dictionary of Old Swedish
and that it has been attested in a number of variant spellings
(fysker, fiisker, fiisk). We also find references to occurrences
of the word in the classical texts, and finally there is a listing
of the compounds in which it occurs, e.g. fiska slagh (Eng.
‘type of fish’). Söderwall (1953) is, basically, intended as
a complement to Söderwall (1884), citing more forms and
more attestations, originating in texts that became available
after Söderwall’s time.2 Schlyter (1887) – as its title indicates – describes the vocabulary of the medieval Swedish
laws, and its entries generally contain a bit less information
than those in Söderwall (1884).
2.
The morphological model used in FM is word and
paradigm, a term coined by Hockett (1954). A paradigm
is a collection of words inflected in the same manner and is
typically illustrated with an inflection table.
An FM lexicon consists of words annotated with
paradigm identifiers from which the inflection engine of
FM computes the full inflection tables.
Consider, for example, the citation form fisker, which
is assigned the paradigm identifier nn_m_fisker. The
paradigm identifier carries no meaning, it could just as well
be any uniquely identifiable symbol, e.g. a number, but we
have chosen a mnemonic encoding. The encoding is read
as: “This is a masculine noun inflected in the same way
as the word fisker” (which is trivially true in this case).
If the paradigm name and the citation form is supplied to
the inflection engine, it would generate the information in
Table 1. To keep the presentation compact, we have contracted some word forms, i.e., the parenthesised letters are
optional.
We also show (in the last column of Table 1) how this
paradigm is presented in traditional grammatical treatises
of Old Swedish, e.g. those by Wessén (1969) and Pettersson (2005). For a discussion of the differences between the
our paradigms and those found in traditional grammatical
descriptions, see section 2.3
The starting point of the paradigmatic specification, besides the dictionaries themselves, are the standard grammars of Old Swedish mentioned above, i.e., those by
Noreen (1904), Wessén (Wessén, 1969; Wessén, 1971;
Wessén, 1965), and Pettersson (2005). The number of
paradigms in the current description by part of speech are
A Computational Morphology
for Old Swedish
2.1. Functional Morphology
The tool we are using to describe the morphological component is Functional Morphology (FM) (Forsberg, 2007;
Forsberg and Ranta, 2004). We chose this tool for a number of reasons: it provides a high-level description language (namely the modern functional programming language Haskell (Jones, 2003; Haskell, 2008)); it uses the
character encoding UTF-8; it supports tasks such as (compound) analysis and synthesis; and, perhaps most importantly, it supports compilation to many standard formats,
such as XML (The World Wide Web Consortium, 2000),
LexC and XFST (Beesley and Karttunen, 2003), GF (Ranta,
2004), and full-form lexicons, and provides facilities for the
user to add new formats.
2
Although Söderwall is given as the author of this work, it was
actually compiled after his death by members of Svenska fornskriftsällskapet (the Swedish Ancient Text Society).
11
Lemma
POS
Gender
Number
sg
sg
sg
sg
pl
pl
nn_m_fisker fisker ⇒
pl
pl
sg
sg
sg
sg
pl
pl
pl
pl
fisker
nn
m
Def
indef
indef
indef
indef
indef
indef
indef
indef
def
def
def
def
def
def
def
def
Traditional
normalized
form
Case
nom
gen
dat
ack
nom
gen
dat
ack
nom
gen
dat
ack
nom
gen
dat
ack
Word form
fisker
fisks
fiski, fiske, fisk
fisk
fiska(r), fiskæ(r)
fiska, fiskæ
fiskum, fiskom
fiska, fiskæ
fiskrin
fisksins
fiskinum, fisk(e)num
fiskin
fiskani(r), fiskæni(r)
fiskanna, fiskænna
fiskumin, fiskomin
fiskana, fiskæna
fiske.r
fisks
fiski, fisk
fisk
fiska(r)
fiska
fiskum
fiska
fiskrin
fisksins
fiskinum
fiskin
fiskani(r)
fiskanna
fiskumin
fiskana
Table 1: The inflection table of fisker
as follows:
Part of speech
Noun
Adjective
Numeral
Pronoun
Adverb
Verb
function that expects a regular noun in nominative singular. If this function is given the word "elephant", then
the result would be another function. This function would,
if an encoding for singular is given to the function, return
{"elephant"}, and if an encoding for plural is given,
return {"elephants"}. The resulting function may be
translated into an inflection table given that the morphosyntactic encoding is ensured to be enumerable and finite (how
this is ensured in FM will not be discussed here).
Turning now to the verb paradigms of Old Swedish,
a Verb is a function from a morphosyntactic encoding,
VerbForm, to a set of word forms with the abstract name
Str.
# of paradigms
38
6
7
15
3
6
2.2. The FM Description
The paradigms of Old Swedish, which in our description
amount to 75 paradigms, are defined using the tool Functional Morphology (FM). We will now give some technical
details of the implementation by explaining how some of
the verb paradigms in our morphology were defined. The
main objective is not to give a complete description, but
rather to provide a taste of what is involved. The interested
reader is referred to one of the FM papers.
An implementation of a new paradigm in FM involves:
a type system; an inflection function for the paradigm; an
interface function that connects the inflection function to
the generic lexicon; and a paradigm name. Note that if the
new paradigm is in a part of speech previously defined, then
no new type system is required.
A paradigm in FM is represented as a function, where
the input is one or more word forms (typically the citation
form or principle parts) and a set of morphosyntactic encodings, and the output of the function is a set of inflected
word forms computed from the input word forms. It is a set
instead of a single word form to enable treatment of variants
and missing cases.
More concretely, if we represent the paradigm of regular nouns in English as a function, and only consider a morphosyntactic encoding for number, we would then define a
type Verb = VerbForm -> Str
The type VerbForm defines the inflectional parameters of Old Swedish verbs. We only include those parameter combinations that actually exist, which will ensure, by
type checking, that no spurious parameter combinations are
created. A morphosyntactic encoding in FM is an algebraic
data type, consisting of a list of constructors, where a constructor may have zero or more arguments. The vertical line
should be interpreted as disjunction. The arguments here
are also algebraic data types (only the definition of Vox is
given here). A member of this type is, for example, Inf
Active, where Active is a constructor of the type Vox.
data VerbForm =
PresSg Modus Vox
PresPl Person Modus Vox
PretInd Number Person Vox
PretConjSg Vox
PretConjPl Person Vox
Inf Vox
ImperSg
ImperPl Person12
12
|
|
|
|
|
|
|
PresSg Ind Pass -> strs [foer++"s"]
Inf v -> passive v [foera]
ImperSg -> strs [foer]
PretInd Pl per v ->
pret_ind_pl (per,v) foer
PretConjPl per v ->
pret_conj_pl (per,v) foer
PretInd Sg _ v -> passive v [foer++"þi"]
PretConjSg v ->
passive v [foer++"þi", foer++"þe"]
_ -> aelska_rule foera p
where foer = tk 1 foera
data Vox =
Active |
Passive
The VerbForm expands into 41 different parameter
combinations. These parameter combinations may be given
any string realization, i.e., we are not stuck with these rather
artificially looking tags, we can choose any tag set. For example, instead of PretConjSg Passive, we have pret
konj sg pass.
The next step is to define some inflection functions. We
start with the paradigm of the first conjugation, exemplified
by the word ælska (Eng. ‘to love’). The inflection function
aelska_rule performs case analysis on the VerbForm
type. There is one input word form, which will be associated with the variable aelska. The function strs translates a list of strings to the abstract type Str. The function
is built up with the support of a set of helper functions,
such as passive that computes the active and passive
forms, tk that removes the nth last characters of a string,
and imperative_pl that computes the plural imperative
forms (inflected for person).
The last inflection function we present, representing
the fourth conjugation paradigm, is liva_rule, defined
in terms of foera_rule. Note that we use two different forms when referring to foera_rule: we use
lif++"a", i.e., the input form where ’v’ has been replaced with ’f’, for the preterite (i.e., past tense) cases, and
the input form for all other cases.
liva_rule :: String -> Verb
liva_rule liva p =
case p of
PresSg Ind Act ->
strs [liv++"er", liv++"ir", liv++"i"]
PresSg Ind Pass -> strs [lif++"s"]
ImperSg
-> strs [lif]
p | is_pret p
-> foera_rule (lif++"a") p
_
-> foera_rule liva p
where liv = tk 1 liva
lif = v_to_f liv
aelska_rule :: String -> Verb
aelska_rule aelska p =
case p of
PresSg Ind Act ->
strs [aelska++"r",aelska]
PresSg Ind Pass -> strs [aelska ++"s"]
Inf v -> passive v [aelska]
ImperSg -> strs [aelska]
ImperPl per ->
imperative_pl per aelsk
PresPl per m v
->
indicative_pl (per,m,v) aelsk
PretInd Pl per v ->
pret_ind_pl (per,v) aelsk
PretConjPl per v ->
pret_conj_pl (per,v) (aelsk++"a")
PresSg Conj v ->
passive v [aelsk++"i",aelsk++"e"]
PretInd Sg _ v -> passive v [aelska++"þi"]
PretConjSg v ->
passive v [aelska++"þi", aelska++"þe"]
where aelsk = tk 1 aelska
When the inflection functions are defined, we continue
with the interface functions. An interface function translates one or more input words, via an inflection function,
into an entry in the generic dictionary. This is done with
the function entry that transforms an inflection function
into an inflection table. If the current part of speech has any
inherent parameters such as gender, those would be added
here. The inherent parameters are not inflectional, they describe properties of a word, which is the reason why they
appear at the entry level.
vb_aelska :: String -> Entry
vb_aelska = entry . aelska_rule
The inflection function aelska_rule computes 65
word forms from one input word form, e.g. kalla (Eng. ‘to
call’).
Given that we now have defined an inflection function for a verb paradigm, we can continue by defining
the other paradigms in relation to this paradigm, i.e.,
we first give the parameter combinations that differ from
aelska_rule and finalize the definition with a reference
to aelska_rule. This is demonstrated in the inflection
function foera_rule, the paradigm of the third conjugation.
vb_foera :: String -> Entry
vb_foera = entry . foera_rule
vb_liva :: String -> Entry
vb_liva = entry . liva_rule
The interface functions need to be named to connect
them with an external lexicon. This is done with the function paradigm. The names are typically the same as those
of the interface functions. Every paradigm is also given a
list of example word forms, which provides paradigm documentation and enables automatic generation of an example inflection table, which is done by FM applying the current interface function to its example word forms. The list
of paradigm names, denoted here with commands, is later
plugged into the generic part of FM.
foera_rule :: String -> Verb
foera_rule foera p =
case p of
PresSg Ind Act ->
strs [foer++"ir", foer++"i"]
13
commands =
paradigm
paradigm
paradigm
no more. We believe that the main factor here is our lack
of information. For many lexical entries it is even difficult
to assign an inflectional class, because the crucial forms
are not attested in the extant texts, and of course, there are
no native speakers on whose linguistic intuitions we could
draw in order to settle the matter.
Some of the diversity built into our paradigms could
thus conceivably be a case of different lexical entries now
brought under the same paradigm, actually consistently using different alternatives for expressing a particular combination of morphosyntactic features; we will probably never
know.
In the standard reference grammars of Old Swedish, inflectional paradigms are consistently idealized in the direction of a (re)constructed Old Swedish, arrived at on the basis of historical-comparative Indo-European and Germanic
studies. In this connection, the actual variation seen in texts
has been interpreted as a sign of language change, of “exceptional” usages, etc.5 (Johnson, 2003). In our paradigms,
we have endeavored to capture the actual variation encountered in the texts and in the dictionary examples (but see
section 3).
[
"vb_aelska" ["ælska"] vb_aelska,
"vb_foera" ["føra"] vb_foera,
"vb_liva"
["liva"] vb_liva
]
We can now start developing our lexicon. The lexicon
consists of a list of words annotated with their respective paradigm, e.g. the word røra (Eng. ‘to touch’) and
føra (Eng. ‘to move’), which is inflected according to the
paradigm vb_foera.
vb_foera "røra" ;
vb_foera "føra" ;
The lines above are put into an external file that is supplied to the compiled runtime system of FM.
2.3.
The Development of the Morphological
Description and the Lexicon
In this project, we have collaborated with a linguist who is
also an expert on orthographic and morphological variation
in Old Swedish. In the first phase of the project, she defined
the inflectional paradigms on the basis of the dictionaries
and the actual variation empirically observed in the texts.
The FM description was developed in parallel with this
work. The linguist selected a set of sample words from
the dictionaries and annotated those with the appropriate
paradigms. The full inflection tables could then be generated immediately and the result evaluated by the linguist.
At the time of writing, about 3,000 main lexical entries
(headwords; see section 1.4) have been provided with inflectional information in the form of a paradigm identifier.
In our work in a parallel project to the one described
here, where we are producing a large computational morphological lexicon for modern Swedish (Borin et al., forthcoming 2008a; Borin et al., forthcoming 2008b), the number of inflectional classes (paradigms) turns out to be on
an order of magnitude more, i.e., around 1,000 rather than
around 100.3 Note that this holds equally for the written
standard language and colloquial spoken Swedish.4 This
is something that calls for an explanation, since under the
(generally accepted, at least in some form) assumption of
uniformitarianism (Janda and Joseph, 2003), we would not
expect to find less diversity in Old Swedish than in the modern language.
First we may note that our morphological description is
not yet complete. For example, while it covers all four weak
verb conjugations, as yet it accounts for only two out of the
nine or so classes of strong and irregular verbs. However,
even standard grammars of Old Swedish like that of Wessén
(1969) list somewhere in the vicinity of 100 paradigms, and
3.
Computational Treatment of Variation
As we have mentioned already (section 1.3), the source of
variation in the texts are of three kinds: no standardized
spelling; no standardized forms; and language change (diachronic drift). For our work on the computational morphological description of Old Swedish, we have found it natural
and useful to make an additional distinction, namely that
between stem variation and ending variation, since it has
seemed to us from the outset that we need to treat stems
and endings differently in this regard.
This gives us altogether six possible combinations of
factors, as shown in the following table:
spelling
variation
lack of lg
standardization
language
change
stem
variation
S1
S2 /L1
S3 /L2
ending
variation
M1
M2
M3
(Legend: S=spelling rules; L=lexical component;
M =morphological component)
5
and possibly even of carelessness or sloth on the part of the
scribes; cf. the following quote, which well captures an attitude
toward linguistic variation traditionally prevalent among linguists:
Variation in Navajo pronunciation had long disturbed
Haile (to Sapir, 30 March 1931: SWL): “Sometimes
I do wish that the informants would be more careful
in pronunciation and follow some system which would
conform to theory. . . . Apparently no excuse, excepting that informants are too lazy to use it correctly.”
Sapir responded (6 April 1931: SWL) that—at least in
collecting texts—it was “not absolutely necessary to
have the same words spelled in exactly the same way
every time.”
(Darnell, 1990, 257)
3
The distribution of inflectional patterns in the modern language is Zipfian in shape: Nearly half the paradigms are singletons, almost a fifth of them have only two members, etc.
4
Although for slightly different reasons in the two cases: In the
written standard language, it is generally the low-frequency words
that have unique paradigms, e.g. learned words and loanwords. In
the spoken language, high-frequency everyday words show variation in their inflectional behavior. There is some overlap, too, e.g.
the strong verbs.
14
reading aid, where it will perform on-the-fly analysis of
words in the texts in order to present the user with possible lexicon entries for the word.
The description is fairly complete, but its usefulness for
this intended practical purpose is still limited by the large
amount of linguistic variation found in the texts.
We have created a small test lexicon (about 3,000 entries), and we are now working on adding inflectional information to all of the headwords in the digital versions of
the Old Swedish reference dictionaries (section 1.4).
We have started to look at the linguistic variation characteristic of the Old Swedish texts. Variation in inflectional
endings is already uniformly handled in the morphological
component, regardless of its origin, while we have still not
made the final decision on a strategy to handle the various
kinds of stem variation found in Old Swedish.
The table reflects the fact that we have decided already to handle all inflectional ending variation – regardless of its origin – in the morphological component, i.e.,
our paradigms contain all attested ending variants, still a finite and in fact rather small set, which partly motivates their
uniform treatment.
Representing a dead language with a finite corpus of
texts, the Old Swedish stems could in theory be treated in
the same way. The corpus is big enough, however, that we
will need to treat it as unlimited in practice, and hence the
stems as a set that cannot be enumerated.
In order not to bite off more than we can chew, we have
tentatively decided to treat all stem variation as a spelling
problem (with one exception; see below). It will then be
natural to look to some kind of solution involving edit distance, e.g. universal Levenshtein automata, see, for example, Mihov and Schulz (2004).
However, the spelling is not completely anarchistic, far
from it: For example, the /i:/ sound will be written <i>, <y>,
<j>, <ii>, <ij>, and possibly in some other ways, but not,
e.g., <a> or <m>, etc. Thus, a rule-based method may be
more appropriate, or possibly a hybrid solution should be
sought.
In the table above, the use of subscripts (S1 , M3 , etc.)
hints at the possibility of distinguishing formally among
different types of information even within a component.
The present morphological description does not make a distinction between ending variation due to spelling variants
of the “same” ending (from a historical-normative point of
view – e.g., indef sg dat fiski/fiske – and ending variation
whereby “different” endings occupy the same paradigmatic
slot, e.g., indef sg dat fiski/fisk. However, there is no technical reason that we could not make this and other distinctions
on the level of paradigms or even on the level of individual
lexical entries. In fact, our work on the Old Swedish morphological description has clearly indicated the need for
this kind of facility.6
There is one kind of stem variation which does not fit
neatly into the picture painted so far, namely that brought
about by inflectional morphological processes, in our case
those of Ablaut and Umlaut. At the moment, the strong verb
class paradigms do not account for variation in the realization of the Ablaut grades of the stem vowels – which of
course we find in the texts – and we are still undecided as
to how to treat them, by a separate normalization step or in
the FM description. In the latter case we would then probably need to duplicate some information already present in
the spelling rules component.
4.
Acknowledgements
The Old Swedish morphological description on which our
computational morphology is based, was made by Rakel
Johnson, Department of Swedish Language, University of
Gothenburg.
The work presented here was financed in part by
Swedish Research Council grant 2005-4211 (2006–2008)
awarded to Aarne Ranta, Chalmers University of Technology, for the research project entitled Library-Based Grammar Engineering, and in part by the Faculty of Arts, University of Gothenburg, through its support to Språkbanken
(the Swedish Language Bank).
We would also like to acknowledge CLT – the Centre
for Language Technology, Göteborg <http://www.clt.gu.
se> – for providing a creative atmosphere in which multidiscplinary collaborations such as this come naturally.
5.
References
Kenneth R. Beesley and Lauri Karttunen. 2003. Finite
State Morphology. CSLI Publications, Stanford University, United States,.
Lars Borin, Markus Forsberg, and Lennart Lönngren.
forthcoming 2008a. The hunting of the BLARK –
SALDO, a freely available lexical database for Swedish
language technology. In Festschrift to Professor Anna
Sågvall Hein. Uppsala University, Dept. of Linguistics
and Philology.
Lars Borin, Markus Forsberg, and Lennart Lönngren.
forthcoming 2008b. SALDO – the Swedish associative thesaurus, version 2. Technical report, Språkbanken,
University of Gothenburg.
Regna Darnell. 1990. Edward Sapir: Linguist, Anthropologist, Humanist. University of California Press, Berkeley
/ Los Angeles / London.
Markus Forsberg and Aarne Ranta. 2004. Functional Morphology. Proceedings of the Ninth ACM SIGPLAN International Conference of Functional Programming, Snowbird, Utah, pages 213–223.
Markus Forsberg. 2007. Three Tools for Language Processing: BNF Converter, Functional Morphology, and
Extract. Ph.D. thesis, Göteborg University and Chalmers
University of Technology.
Summary and Conclusions
We have implemented a morphological description for Old
Swedish using Functional Morphology, a tool which supports automatic morphological analysis and generation, as
well as the generation of full-form lexicons. The description is intended to be used in an online Old Swedish text
6
It is not difficult to think of situations where this would be
useful in modern language descriptions as well; for instance, it
would be useful to be able to record the frequency of occurrence
of homographs according to which lexical entry they represent.
15
Haskell. 2008. Haskell homepage. <http://www.haskell.
org>.
Charles Hockett. 1954. Two models of grammatical description. Word, 10:210–234.
Richard D. Janda and Brian D. Joseph. 2003. On language,
change, and language change – or on history, linguistics, and historical linguistics. In Brian D. Joseph and
Richard D. Janda, editors, Handbook of Historical Linguistics, pages 3–180. Blackwell, Oxford.
Rakel Johnson. 2003. Skrivaren och språket. Ph.D. thesis,
Department of Swedish Language, Göteborg University.
Simon Peyton Jones. 2003. Haskell 98 Language and
Libraries: The Revised Report. Cambridge University
Press.
Fred Karlsson, Atro Voutilainen, Juha Heikkilä, and
Arto Anttila, editors. 1995. Constraint Grammar: A
Language-Independent System for Parsing Unrestricted
Text. Number 4 in Natural Language Processing. Mouton de Gruyter, Berlin and New York.
Stoyan Mihov and Klaus Schulz. 2004. Fast approximate
search in large dictionaries. Computational Linguistics,
30(4):451–477.
John Nerbonne and Petra Smit. 1996. GLOSSER-RuG:
In support of reading. In COLING–96. The 16th International Conference on Computational Linguistics. Proceedings, Vol. 2, pages 830–835, Copenhagen. ACL.
Adolf Noreen. 1904. Altschwedische Grammatik. Halle.
Facsimile available online: <http://lexicon.ff.cuni.cz/
texts/oswed_noreen_about.html>.
Gertrud Pettersson. 2005. Svenska språket under sjuhundra år. Studentlitteratur, Lund, Sweden.
Aarne Ranta. 2004. Grammatical Framework: A Typetheoretical Grammar Formalism. The Journal of Functional Programming, 14(2):145–189.
P. Rayson, D. Archer, A. Baron, J. Culpeper, and N. Smith.
2007. Tagging the Bard: Evaluating the accuracy of a
modern POS tagger on Early Modern English corpora.
In Proceedings of Corpus Linguistics 2007. University
of Birmingham, UK.
C.J. Schlyter. 1887. Ordbok till Samlingen af Sweriges
Gamla Lagar. (Saml. af Sweriges Gamla Lagar 13).
Lund, Sweden.
Knut Fredrik Söderwall. 1884. Ordbok Öfver svenska
medeltids-språket. Vol I–III. Lund, Sweden.
Knut Fredrik Söderwall. 1953. Ordbok Öfver svenska
medeltids-språket. Supplement. Vol IV–V. Lund, Sweden.
The World Wide Web Consortium. 2000. Extensible Markup Language (XML). <http://www.w3.org/
XML/>.
Elias Wessén. 1965. Svensk språkhistoria: Grundlinjer till
en historisk syntax. Stockholm, Sweden.
Elias Wessén. 1969. Svensk språkhistoria: Ljudlära och
ordböjningslära. Stockholm, Sweden.
Elias Wessén. 1971. Svensk språkhistoria: Ordböjningslära. Stockholm, Sweden.
16
Generating Tailored Texts for Museum Exhibits
Dana Dannélls
Natural Language Processing Research Unit
Department of Swedish Language
University of Gothenburg
SE-405 30 Gothenburg, Sweden
[email protected]
Abstract
This paper reports work that aims to generate texts in multiple languages from ontologies following the Conceptual Reference Model
(CRM) ISO standard for conceptual models of museums. The rationale of this work is to increase users’ knowledge and interest in
the cultural heritage domain by allowing the user to select his preferable syntax presentation and influence the order of the generated
information using generation techniques and Semantic Web technologies. We chose for study a small amount of logical relations represented in the ontology and wrote a grammar that is capable to describe them in natural language through user editing. We present the
multilingual source authoring environment, which is built upon the grammatical framework (GF) formalism and show how it is utilized
to generate multiple texts from the CRM domain ontology. The initial results comprise texts, which vary in syntax and content.
1.
Introduction
manipulate the content of the document indirectly in his/her
own language. An example of a template-based authoring
tool that makes use of this technique within the CH domain
was presented by Androutsopoulos et al. (2007). An alternative approach to template-based NLG that is particularly
relevant in cases where texts are generated from logical
forms in several languages simultaneously is a grammarbased approach (Bateman, 1997).
In this paper we present a multilingual source authoring
tool, which is built upon the grammatical framework (GF)
formalism to generate texts from the underlying semantic representation that is based on the Conceptual Reference Model (CRM) domain ontology. The authoring environment is similar to those described in Power and Scott
(1998), Dymetman et al. (2000) and van Deemter and
Power (2003). 2 The focus is on the process starting from a
fixed semantic representation to a surface realization, with
emphasis on the syntactical sentence structure, and the content variation.
The structure of the paper is as follows. In section 2 we
elaborate the notion of ontology and describe both the reference ontology model and the grammar formalism that our
application is built upon. Section 3 presents the grammar
implementation and explains how it is utilized to generate tailored descriptions from a formal representation language. Section 4 finishes with conclusions and a discussion
of future work.
During the last decade, the awareness of the need for personalization has become fundamental for cultural institutions such as museums and libraries while aim to produce
textual descriptions of museum exhibits tailored to the visitor’s knowledge, interests, and personal preferences, such
as preferred vocabulary, syntax, sentence length etc. One
of the first examples of personalization in a museum context was developed in the Intelligent Labelling Explorer
(ILEX) project,1 by using Natural Language Generation
(NLG) techniques. More recently, applications within the
cultural heritage (CH) domain have seen an explosion of
interest in these techniques (Novello and Callaway, 2003;
O’Donnell et al., 2001; Androutsopoulos et al., 2007).
The process of NLG starts from an ontology that describes
a certain domain. Recently, natural language generators
that are targeted towards the Semantic Web ontologies have
started to emerge. A strong motivation for generating texts
from ontologies is that the information represented in an
ontology has a true potential to provide a large amount of
text if this text is realized correctly. Gradually, the cultural
heritage knowledge domain which is often characterized by
complex semantic structures and large amounts of information from several different sources will benefit from the
complete generation of the information delivered in the ontology.
Web ontology languages pose many opportunities and challenges for language generators. Although standards for
specifying ontologies provide common representations to
generate from, existing generation components are not
compatible with the requirements posed by these newcoming standards. This issue has been previously addressed by developing domain-dependent authoring interfaces that are built upon an ontology and that allows it to
be deployed through knowledge editing (Brun et al., 2000;
Hartley et al., 2001; van Deemter et al., 2005). These interfaces are links between the ontology and the user who can
1
2.
Background
In the context of the work presented here, an ontology is
understood as a formal model that allows reasoning about
concepts, objects and about the complex relation between
them. An ontology holds meta-level information about different types of entities in a certain domain and provides a
structure for representing contexts, it is not human readable
as it is designed to be processed by computer systems.
2
http://www.hcrc.ed.ac.uk/Site/ILEXINTE.html
17
The advantages of utilizing this family of domain authoring
approaches that are coupled with multilingual text generation are
elaborated in Scott (1999).
2.3. The Grammatical Framework (GF)
The Grammatical Framework (Ranta, 2004) is a functional grammar formalism based on Martin-Löf’s typetheory (Martin-Löf, 1973) implemented in Haskell.6 GF
focuses on language independent semantic representations.
It differentiates between domain dependent and domain independent linguistic resources, as it is designed to be applicable both to natural and to formal languages. One abstract
grammar can have several corresponding concrete grammars; a concrete grammar specifies how the abstract grammar rules should be linearized in a compositional manner.
Multilingual functional grammatical descriptions permit
the grammar to be specified at a variety of levels of abstraction, which is especially relevant for constructing a detailed mapping from semantics to form. This aspect is crucial for natural language generation to work. What makes
the grammar suitable for generating from ontologies and in
particular from OWL, is that it allows multiple inheritance.
GF has three main module types: abstract, concrete, and
resource. Abstract and concrete modules are top-level, in
the sense that they appear in grammars that are used at runtime for parsing and generation. They can be organized into
inheritance hierarchies in the same way as object-oriented
programs. The main advantage with converting the ontology to GF is that we can make use of the rich type system in
the concrete syntax for capturing morphological variations.
Our approach is based on the idea suggested by Khegai et
al. (2003) who utilized GF to automatically generate multiple texts from semantic representations. The source authoring environment deploys similar techniques to those introduced in Power and Scott (1998), Dymetman et al. (2000)
and van Deemter and Power (2003).
Examples of Web ontology-languages that have been developed by the W3C Web-Ontology working group are OWL
and DAML+OIL.3 The basis for the design of these Web
technology languages based on the RDF Schema is the expressive Description Logic (DL) SHIQ (Horrocks et al.,
2003). These languages provide extensive reasoning capabilities about concepts, objects and relationships between
them.
2.1.
Generating from an Ontology
In an ontology, an object may be described by semantic
graphs whose nodes (concepts) represent parts of an object,
and the arcs (relations) represent partial constrains between
object parts. Each relation described in a logical language
is binary, i.e. it connects between two nodes. In order to
present a piece of information about an object represented
in an ontology, multiple sentences must be formulated. It
becomes valuable if these sentences that build the final text
can be adapted to various contexts or users.
There has been successful attempts to generate from ontologies (Wilcock, 2003; Wilcock and Jokinen, 2003;
Bontcheva and Wilks, 2004; Bontcheva, 2005). Wilcock
(2003) and Wilcock and Jokinen (2003) have shown how
RDF/XML generation approach can be extended so that
the information embedded in the ontology can be exploited
to generate texts from Web ontology-languages such as
DAML+OIL and OWL without the need for a lexicon.
Bontcheva (2005) demonstrated how to minimize the effort when generating from Web ontology-languages while
being more flexible than ontology verbalisers. Some of the
difficulties reported by these authors concern lexicalization
and in establishing context variations.
3.
2.2.
The CRM Ontology
Generating from the Ontology
We chose for study a small amount of logical relations
represented in the ontology and wrote a grammar that is
capable to describe them in natural language through user
editing. The following code is a fragment taken from the
ontology we employed. The code states that the class
PaintingP9091 must have at least one value TypeValue on
property has type; the individual TypeValue is an instance
of the class cidoc:E55.Type7 and has two property values:
“tool” and “painting”.
One initiative to enable an ontology in the context of the
cultural heritage is the Conceptual Reference Model domain ontology. The International Committee for Documentation of the International Council of Museums Conceptual
Reference Model (CIDOC-CRM)4 is a core ontology and
ISO standard for the semantic integration of cultural information with library archive and other information (Doerr,
2005). The primary role of the CRM is to enable information exchange and integration between heterogeneous
sources of cultural heritage information.
The central idea of the CIDOC-CRM is that the notion of
historical context can be abstracted as things and people.
It concentrates on the definition of relationships rather than
classes to capture the underlying semantics of multiple data
and meta structures. It tends to provide an optimal analysis of the intellectual structure of cultural documentation in
logical terms, which is available in several formats such as
RDF and OWL that have hardly been explored yet. The
work described in this paper is based on the OWL version
of the ontology.5
<owl:Class rdf:about=”PaintingP9091”>
<owl:Restriction>
<owl:onProperty rdf:resource=”&cidoc;P2F.has type”/>
<owl:hasValue rdf:resource=”#TypeValue”/>
</owl:Restriction>
</owl:Class>
<owl:Thing rdf:about=”#TypeValue”>
< rdf:type rdf:resource=”&cidoc;E55.Type”/>
<Tool rdf:datatype=”&xsd;string”>tool
</Tool>
<Painting rdf:datatype=”&xsd;string”>painting
6
Haskell is a standardized purely functional programming language with non-strict semantics. Similar to Lisp and Scheme.
7
The notation &cidoc; is used instead of the whole namespace,i.e http://cidoc.ics.forth.gr/OWL/cidoc v4.2.owl#
3
http://www.w3.org/TR/
4
http://cidoc.ics.forth.gr/
5
http://cidoc.ics.forth.gr/OWL/cidoc v4.2.owl
18
3.2. The Concrete Representation
Each category and function introduced in the abstract syntax has a corresponding linearization type in the concrete
syntax. Linearization rules are declared differently for
each target language. In addition, each concrete syntax
also contains grammatical parameters and grammar rules,
which are used to ensure grammatical correctness for each
language, in our case English and Swedish. An example of
linearization rules taken from the English concrete syntax
is the following:
</Painting>
</owl:Thing>
The above fragment exemplifies the representation of the
classes and relationships that are utilized by the grammar.
In the grammar implementation, classes are represented
as categories; properties are functions (rules) between two
categories, where each property links between two classes;
individuals are lexical categories (strings). Below is a
representation of the mkObject, which corresponds to a
function that links between the classes of an Object:
lin
mkObject:ObjectNodeI→ObjectNodeII→ObjectNodeIII→Object;
CarriedOutBy Painting obj = {s = det ! obj.num
++ cop ! obj.num ++ “painted by” ++ obj.s ;
num=obj.num};
In this example the Object category corresponds to
PaintingP9091. Each ObjectNode is a class, according to the above ontology representation, ObjectNodeI
corresponds to the cidoc class cidoc:E55.Type. It is
followed by ObjectNodeII, i.e. cidoc:E52.Time-Span and
ObjectNodeIII, i.e. cidoc:E21.Person, as shown below.
Painting = {s = “painting” ; num = sg} ;
Painting = {s = “paintings” ; num = pl} ;
Grammatical features are supported by GF and the agreement between the pronoun and the verb is enforced in the
generated sentences. The variable obj represents a terminal string. The parameter num is an abbreviation for the
parameter type “number”, it contains the inherent number
that can be either singular (sg) or plural (pl). The operation
det is a determiner, and the operation cop is copula verb.
{Type} instance of ObjectNodeI
{Time-Span} instance of ObjectNodeII
{Person} instance of ObjectNodeIII
Consequently, individuals such as “tool” and “painting” are
terminals and are declared in the concrete syntax. In the
next sections we describe the abstract and the concrete representations
3.1.
3.3.
The Authoring Environment
Figure 3 illustrates the source authoring environment. The
left-side window shows the abstract syntax tree, which represents the Object structure. The large window positioned
to the right is the linearization area, the editing focus is
presented as the highlighted metavariable ?3. The bottom
area shows the context-dependent refinement for the ObjecNodeIII, there are two possible relations to choose from.
The Abstract Representation
The abstract syntax is a context-free grammar where each
rule has a unique name. An abstract rule in GF is written
as a typed function. The categories and functions are
specified in GF by cat and fun declarations. Below is a
fragment of the grammar:
cat
Object ;ObjectNodeI ; Type ;
ObjectNodeII ; Time-Span ;
ObjectNodeIII ; Person ;
fun
HasType This : Type → ObjectNodeI;
HasType Here : Type → ObjectNodeI;
HasType Template : Type → ObjectNodeI;
HasTimeSpan: Time-Span → ObjectNodeII;
CarriedOutBy Painting: Person → ObjectNodeIII;
CarriedOutBy Tool: Person → ObjectNodeIII;
The abstract syntax gives a structural description of a part
of the domain. It has several advantages, one of which
is the ability to utilize the same categories differently depending on the semantic complexity of the context. Here
we declared three functions for the ObjectNodeI to achieve
context variations, though very simple ones. Similarly, we
declared two functions for the ObjectNodeIII, however, the
difference between CarriedOutBy Painting and CarriedOutBy Tool is the choice of the verb in the linearization
rule. The verb painted by is applied when the subject is
the noun painting, but the verb created by is applied when
the subject is the noun tool, in cases when the object is an
instance that belongs to the category Person.
Figure 1: The GF source authoring environment.
The authoring tool that is built upon the GF grammar makes
it possible to generate the following texts:
English
(1) Here we have a painting. It was painted by
Carl-Johan Harvy. It was made in 1880.
(2) This is a tool. It was made in 1880. It was
created by Carl-Johan Harvy.
19
(3) On the second floor of the history museum
we have paintings. They were created by Siri
Derkert. They were produced in Italy.
K. Bontcheva. 2005. Generating tailored textual summaries from ontologies. In Second European Semantic
Web Conference (ESWC’05), Crete.
C. Brun, M. Dymetman, and V. Lux. 2000. Document
structure and multilingual authoring. In In Proc. of First
International Natural Language Generation Conference
(INLG), Mitzpe Ramon, Israel, June.
M. Doerr. 2005. The cidoc crm, an ontological approach
to scheme heterogeneity. In Dagstuhl Seminar, editor,
Semantic Interoperability and Integration, pages 1862–
4405.
M. Dymetman, V. Lux, and A. Ranta. 2000. Xml and multilingual document authoring: Convergent trends. In In
Proceedings of COLING. Saarbrucken, Germany.
A. Hartley, D. Scott, J. Bateman, and D. Dochev. 2001.
Agile – a system for multilingual generation of technical
instructions. In In MT Summit VIII, Machine Translation
in the Information Age, pages 145–150.
I. Horrocks, P.F. Patel-Schneider, and F. van Harmelen.
2003. From shiq and rdf to owl: The making of a web
ontology language. Journal of Web Semantics, 1(1):7–
26.
J. Khegai, B. Nordström, and A. Ranta. 2003. Multilingual
syntax editing in gf. pages 453–464, Mexico. Springer.
P. Martin-Löf. 1973. An intuitionistic theory of types:
Predicative part. In H. E. Rose and J. C. Shepherdson,
editors, In Proc. of Logic Colloquium ’73, Bristol, UK,
volume 80, pages 73–118.
A. Novello and C. Callaway. 2003. Porting to an Italian
surface realizer: A case study. In Proc. of the 9th European Workshop on NLG, pages 71–78.
M. O’Donnell, J. Oberlander C. Mellish, and A. Knott.
2001. Ilex: an architecture for a dynamic hypertext generation system. Natural Language Engineering,
7(3):225–250.
R. Power and D. Scott. 1998. Multilingual authoring using feedback texts. In 17th International Conference on
Computational Linguistics, pages 1053–1059.
A. Ranta. 2004. Grammatical framework, a typetheoretical grammar formalism. Journal of Functional
Programming, 14(2):145–189.
D. Scott. 1999. The multilingual generation game: Authoring fluent texts in unfamiliar languages. In Proc. of the
16th International Joint Conference on Artificial Intelligence (IJCAI), pages 1407–1411.
K. van Deemter and R. Power. 2003. High-level authoring
of illustrated documents. Natural Language Engineering., 9(2):101–126.
K. van Deemter, E. Krahmer, and M. Theune. 2005. Real
versus template-based natural language generation: a
false opposition? Computational Linguistics, 31(1):15–
23.
G. Wilcock and K. Jokinen. 2003. Generating responses
and explanations from rdf/xml and daml+oil. In Knowledge and Reasoning in Practical Dialogue Systems
IJCAI-2003, pages 58–63.
G. Wilcock. 2003. Talking owls: Towards an ontology verbalizer. In Human Language Technology for the Semantic Web and Web Services, pages 109–112.
Swedish
(1) Här har vi en målning. Den är målad av
Carl-Johan Harvy. Den är gjord på 1880 talet.
(2) Det här är ett redskap. Det är gjort på 1880
talet. Det är tillverkat av Carl-Johan Harvy.
(3) På andra våningen i historiska museet har vi
målningar. De är tillverkade av Siri Derkert. De
är producerade i Italien.
The difference between the first and second sentence is the
order in which the ObjectNodeII and the ObjectNodeIII appears, this is done with the help of the variants function that
allows for syntactic variations by reordering the linearized
categories. The third sentence illustrates a typical example of a combined template and grammar based generation,
e.g. the fixed sentence: “On the second floor of the history
museum” that has been prewritten.
4.
Conclusions and Future Work
In this paper we present a multilingual grammar-based approach, the aim of which is to generate exhibit descriptions
following the CRM domain ontology. We present the GF
authoring tool, which allows users to choose the content
and the form of the output text. In this work we attempt
to establish new methods that support the user on receiving information based on the Semantic Web in the cultural
heritage domain.
We chose for study a small amount of logical relations represented in an ontology and have started to examine the capabilities of utilizing a grammar to bridge between ontology representations and different users. This is one of the
few attempts to generate texts from CIDOC CRM.
Future work will focus on ontology studies and on particular problems of generating for cultural heritage. We are
also planning to utilize the Resource Grammar Library that
has been developed to provide the linguistic details for application grammars on different domains. This will be a
step towards high quality summary generation. Our goal is
to built a grammar that reflects the ontology structure and
supports all the OWL features to allow the user to interact
with the full ontology.
5.
References
I. Androutsopoulos, J. Oberlander, and V. Karkaletsis.
2007. Source authoring for multilingual generation of
personalised object descriptions. Natural Language Engineering, 13(3):191–233.
J. A. Bateman. 1997. Enabling technology for multilingual
natural language generation: The kpml development environment. Natural Language Engineering, 3(15–55).
K. Bontcheva and Y. Wilks. 2004. Automatic report generation from ontologies: the miakt approach. In Nineth International Conference on Applications of Natural Language to Information Systems (NLDB). Manchester.
20
Non-Standard Russian in Russian National Corpus (RNC)
Elena Grishina
Institute of Russian Language, RAS
121019, Volkhonka, 18/2, Moscow, Russia
[email protected]
Abstract
The RNC is a 150 million-word collection of Russian texts, thus, it is the most representative and authoritative corpus of the
Russian language. It is available on the Internet at www.ruscorpora.ru. The RNC contains texts of all genres and types, which
cover Russian of the 18-21 centuries. That’s why the problem of non-standard units in the RNC and the possible ways of their
normalization is now the most important and urgent. To solve the problem we suggest 1) expanding the base of RNC Vocabulary, 2) improving the RNC Grammar parser in some dimensions, 3) using the preserving annotation in some kinds of texts.
The paper describes what type of non-standard units may be normalized with this or that strategy of standardizing and which
are the cases, when the users paradoxically are not interested in transforming the non-standard units into the standard ones. The
paper contains considerable number of non-standard samples from the sub-corpora of the RNC – the Spoken, Electronic, Accentological Corpora, the Corpus of Old Russian Texts and of Poetry.
•
the Corpus of Poetry (Russian poetry of the 18–21
centuries, 1,9 million tokens)
• the Corpus of Spoken Russian (the scripts of spoken texts and movie transcripts of the period of
1934–2007, 5,5 million tokens)
• the Corpus of Dialect Texts (contains the records
of dialect specimens from different regions of Russia, 0,144 million tokens)
• the Accentological Corpus (contains Russian poetry with marked arses2 and the movie transcripts
with marked stress; the Accentological Corpus will
be available at 2008).
It’s quite natural that any corpus of so large capacity and
of so wide spectrum of text databases as the RNC contains
a lot of non-standard words and wordforms. That’s why
the creators of the RNC were forced to elaborate the integral and consecutive strategy to settle the Non-Standard
Units (NSU) and to turn them into the standards, as far as
possible.
1. Russian National Corpus:
Content and Structure
The Russian National Corpus (RNC) is a collection of
Russian texts of the 18–21 centuries, which are supplied
with different types of annotation – morphological, semantic, accentological, sociological, metatextual.
The RNC has been functioning since April 2003 and is
accessible at www.ruscorpora.ru. Now it contains circa
150 million words. It’s planned to bring its capacity up to
200 million words1. Therefore, the today RNC has become one of the most representative and authoritative
corpora of Russian.
The RNC embraces all genres of Russian written texts –
fiction (prose, drama, poetry) and non-fiction (journalism,
scientific texts, memoirs, letters, business documents,
theological writings, etc.). Also the RNC contains the
scripts of Spoken Russian, including movie transcripts. A
user of the RNC can form his sub-corpus according to any
parameter of metatextual, sociological, morphological and
semantic annotation and to any combination of all possible parameters.
The metatextual annotation describes a text as a whole
(its title, author, date of creation, genre/type, and so on).
The morphological annotation assigns the grammatical
characteristics to a wordform. The semantic annotation
characterizes a lexeme as a member of this or that semantic class, e. g. names of persons, of animals, diminutives,
adjectives of good-rating, verbs of speech or moving, and
so on. The sociological annotation is used within the Corpus of Spoken Russian (see below): every token is annotated from the point of view of sex and age of a speaker (if
they are known), and thus a user may form his own subcorpus).
The whole structure of the RNC now looks as follows:
•
•
2. Types of NSU and Technology
of Normalization
Any token in the RNC is characterized as non-standard if
it isn’t presented in the RNC Vocabulary. The Vocabulary
consists of three zones: 1) Lemma (the title of the entry,
the citation form of a word), 2) Wordform (the set of the
Inflected Forms of a Lemma), 3) Grammar (grammatical
characteristics of an inflected form). Therefore, NSU may
be non-standard as a lemma, as a wordform, or from the
point of view of its grammatical properties.
These types of NSU are to be exemplified.
2.1.
Lemma-NSU
Obviously, this case is the most frequent, and exactly
Lemma-NSU occur when we’re speaking about nonstandard units.
the Main Corpus [97,5 million tokens] – the corpus
of contemporary Russian (the period between the
middle of the 20 century and the beginning of the
21 century; the Main Corpus also includes the
Electronic Sub-Corpus [1,2 million tokens] – the
collection of the internet texts)
2.1.1. Proper Names
The proper names form the extensive class of linguistic
units, which have some specific grammatical characteristic. The class includes toponyms and anthroponyms (first
names, including pet names, family names, patronymics).
the Corpus of Early Texts (the 18 – the middle of
the 20 century, 51 million tokens)
2.1.2. Abbreviations
The class includes acronyms and initialisms.
1
2
The prehistory and the contemporary state of the RNC are
available to be learnt about in (NKRJA, 2005, 2008).
‘Arsis’ (plural ‘arses’) is the special term in classical prosody
for a syllable on which the ictus falls in a metrical foot.
21
The process of the expanding consists not only of the determination of the list of the supplementary lemmas, but
also of their grammaticalization: we ought to adjust a
lemma to its grammatical characteristics (the grammatical
index, which describes a lemma’s type of paradigm and
other grammatical parameters).
Naturally, the main body of the grammaticalization of the
new entries is carried out with the grammar parser3. But
the grammatical structure of Russian doesn’t permit us to
rely totally on the work of the parser: we need to check
the results carefully by ourselves.
For example, the parser never knows, how to mark the
name of the town Suzdal’: it may be masculine or feminine, and this difference leads to the difference in the case
endings. Therefore, only a linguist can choose the right
gender (masculine, in this case). Naturally, the toponym
Suzdal’ is quite frequent and well-known, so it is included
in all Russian toponymic dictionaries. But we face far
more difficult cases, e. g. the town name Ivdel’, which is
usually masculine, but the natives very often decline it as
feminine. The toponymic dictionaries miss this place
name, therefore, here we need a linguist to solve the problem.
Only a linguist can distinguish the family name Kurchatov
from the toponym Kurchatov: their paradigms differ in
several position (for example, InstrSg Kurchatovym vs.
Kurchatovom).
Only in manual mode the grammar characteristics of new
adoptions can be defined: for example, praimeriz (< Eng.
primaries) or g’orlz (< Eng. girls): the parser automatically marks them like ‘masculine Singular, o-declension’,
when really they are ‘Plural indeclinable’.
A priori it is not known, what are the numeral and gender
characteristics of an abbreviation. For example, the initialisms of the similar literal structure may have different set
of grammatical characteristics: CRU = SingNeu, SBU =
SingFem, SŠA = Plur, etc.
2.1.3. New Loanwords
Russian is a language which is quite open to adoptions, so
the stream of loanwords never subsides.
2.1.4. Obsolete and Rare Words
As it was mentioned above, the RNC contains the Corpus
of Early Texts and the Corpus of Poetry, so it is quite
natural that the quantity of dated, rare and stylistically
marked words in the RNC is considerable.
2.1.5. Substandard Vocabulary
On the other hand, provided that the RNC includes a lot of
patterns of modern literature, the Sub-Corpus of Electronic Communication, the patterns of private Spoken
Russian, and the great amount of journalism, the presence
of the substandard words and word combinations in the
Corpus is of no wonder.
2.1.6. How We Solve the Problems
The only way to deal properly with the Lemma-NSU,
listed in the §§2.1.1–2.1.5, is to expand the RNC Vocabulary regularly. The sources of the expanding may be as
follows.
• Existing dictionaries of Russian. Substantially, the
problems, named in §§2.1.1, 2.1.3–2.1.5, were
solved by means of some Russian dictionaries –
dictionaries of proper names (Petrovsky, 1996;
Gramm (supplement), 2003), of loanwords (SIS,
2008), of substandard words (Elistratov, 2000).
The word-lists of the fundamental dictionaries of
Russian (Ortho, 2005; Gramm, 2003) also were of
use. Naturally, it would be more convenient to use
for these purposes the Russian Ontologies, as it is
often done when NERC (Named-Entity Recognition and Classification) is used (see (Nadeau, 2007)
among others). But the fact is that the generation of
the Russian ontologies is now at its starting stage
(see, e. g. (Nevzorova at al., 2007), and, above all,
the RNC itself may be used as one of the source of
the required ontology.
• The list of wordforms, which lack grammar. It is
widely known, that the language development always surpasses the dictionaries development.
Therefore, the Vocabulary of the RNC itself is the
reliable source for expanding of the RNC: we ought
to make the RNC profile after each completion of
the Corpus and then to expand the RNC Vocabulary. For example, contemporary Russian lacks a
dictionary of abbreviations, so the only way to
solve the problem of abbreviations (see §2.1.2) is to
address to the RNC itself.
• The regular derivation. It takes place when one of
the groups of Lemma-NSU are the derivatives of
the other, meaning that the second group is generated from the first one according to the set of regular rules. For example, Russian patronymics are the
derivatives of the Russian names, and the list of the
derivative rules is relatively short and clear. Therefore, to expand the list of patronymics in the RNC
Vocabulary (see §2.1.1), there is no need to refer to
dictionaries: we can receive the necessary units, using the dictionary of the Russian first names (for
example, (Petrovsky, 1996; Gramm (supplement),
2003)) and the set of derivation programs.
2.1.7. Graphic Corruptions
The Corpus of Electronic Communication, the Corpus of
Early Texts and the Corpus of Poetry raise the question of
the graphic and orthographical corruptions.
1) The text may be corrupted orthographically, when
an author reflects in his spelling the way of his own
pronunciation or when there is a misprint.
The simple way to solve the problem of orthographical
corruption is to use automatic speller, when a text is being
prepared to be included in the RNC. The programs of the
kind may automatically change a wrong spelling into a
right one (when we deal with the unambiguous correlation
between the wrong spelling and the right spelling)4.
But the fact is that there is a particular class of the RNC
users, who doesn’t want us to correct wrong spellings.
Firstly, these are the specialists in Russian orthography.
They investigate the contemporary tendencies in Russian
3
The RNC uses two grammar parsers: DiaLing (in the Disambiguated Sub-Corpus) and MyStem (in the Main Corpus). Both
parsers are PoS-taggers, not syntactical ones: the fact is that the
morphology of Russian gives a user the possibility to solve a lot
of the problems of Russian syntax.
4
Therefore, nowadays we may mirror the early days situation:
formerly, a corpus was the testing area for a speller (see, among
others, (Chae Young-Soog, 1998), today, vice-versa, we may use
a speller to improve a corpus.
22
As the examples of the netspeak graphical corruptions we
may consider so called yazyk padonkaff ‘the language of
scums’, or, in other terms, olbanskiy yazyk ‘the Albanian
language’. It’s mainly used in Internet, and its main characteristics is the deliberate distortion of the literal cover of
a word and the ostentatious contrast to the common orthography. For example: the spelling padonkaff instead of
the normative podonkov, the spelling olbanskiy instead of
albaskiy, and so on.
The lazy reason to use graphical corruptions is to save
writer’s efforts and time (see, for example, regular way of
writing sobsno instead of sobstvenno ‘actually’, scha
(Russian ща) instead of seychas ‘now’, and 2 (to) and 4
(for) in English.
In Russian Internet communication the enciphering
graphical corruptions is widely used. It’s due to the strong
policy of moderating and to the implied automatic elimination of taboo words. It is widely known, that Russian is
a language with the ramified system of taboo-words,
which are used with pleasure by the visitors of the Internet
forums. So, to trick an automatic moderator, the visitors
use the enciphered corruptions. For example, a) the spelling 3,14zdet’ instead of pizdet’ [taboo for ‘talk nonsense,
gabble’, the word itself is included in the RNC Vocabulary], because 3,14 is the value of pi(π); b) the spelling
2,72bat’ instead of ebat’ [taboo for ‘make love with smb’,
the word itself is included in the RNC Vocabulary], because 2,72 is the value of e, the base of natural logarithm;
c) the spelling @bat’ (see previous point), because @ is
the symbol of e-mail. And so on.
Due to the small number of graphical corruptions and
because of their fluctuation and instability we can’t treat
them automatically. Moreover, we mustn’t add them to
the RNC Vocabulary. So, the only way to transform all
these NSU is the preserving annotation (see above),
started up in manual mode. Such mode is widely used to
edit the texts of Internet communication to expand the
RNC.
spelling to elaborate timely recommendations for the
change of the set of the Russian orthographical rules.
Naturally, they want to know, what is the real orthographical practice. On the other hand, the investigators of
the old Russian texts (for example, the texts of the 18 century) are specially interested in orthographical practice of
the period. Therefore, their scientific aspirations would be
violated, if the creators of the RNC automatically correct
the old orthography and make it standard.
Therefore, in this case we have the only way to solve the
problem: an inaccurate spelling ought to be preserved in
order that any concerned user can find it in a text. On the
other hand, that very incorrectness must be supplied with
the correct form. This process may be accomplished
automatically, with the help of speller, or in manual mode.
Anyway, as a final result we receive an annotation of the
following structure:
Incorrectness{CorrectSpelling + Grammar Characteristics + Semantic
Characteristic}.
Inc is a usual unit of the text, so it may respond to a usual
query, as any other word in the Corpus; CorSp is the virtual correct unit, which corresponds to Inc, and also may
respond to a usual query (but CorSp is invisible for a user,
so it’s presence in a text is virtual); Gram and Sem are the
grammatical and the semantic characteristics of the CorSp
respectively.
This way to annotate NSU is called preserving annotation. The preserving annotation gives us the following
possibilities: 1) to find all contexts, which contain that
very Inc; 2) to find all possible contexts, which contain
that very CorSp, including the contexts with corresponding Inc; 3) to find all possible contexts, containing CorSp,
which shows itself as Inc; 4) to find all possible contexts,
which contain that very Gramm, including the contexts
with Inc, the contexts with CorSp, or both; 5) to find all
possible contexts, which contain that very Sem, including
the contexts with Inc, the contexts with CorSp, or both.
Naturally, the preserving annotation is used in manual
mode when the annotated texts are of exclusive importance (for example, some texts of 18 century, of electronic
communication, or scripts of Spoken Russian). Also the
preserving annotation is used, when we deal with the
Educational Sub-Corpus of RNC (designed for school
teaching) and with the Disambiguated Sub-Corpus (its
volume is about 5 million tokens), where the grammatical
homonymy was disambiguated in manual mode.
In other cases we may use the orthoparser, which would
1) mark Inc5 in analyzed text, 2) automatically generate
the set of corresponding CorSp, and 3) assign to CorSp
necessary Gramm and Sem. For the Russian this problem
may be solved on the basis of the existing spellers at an
early date.
2) When the author deliberately enciphers his text, we
deal with graphical corruptions. The reason for
graphical corruptions are multifarious, but the main
ones are as follows: 1) netspeak, 2) lazy, and 3) enciphering proper.
2.1.8. Word-Forming Doublets
Some derivative units in Russian are duplicated with
slight differences in phonetics, grammar and stylistics.
This doublets always have the same etymological history.
The dictionaries of Russian never include all these doublets – in most cases they contain only one of the twins, so
the other is treated in the dictionaries as a bastard and
therefore is never marked and annotated in the RNC.
2.1.9. The Open Sets of Derivates
This class of linguistic units is quite a problem for any
dictionary. The RNC Vocabulary is no exception. All the
derivatives, which are created according to any productive
word-formative model, can’t be included in a dictionary.
The lexicographers usually focus on the frequency of this
or that derivative, and take their decision, concerning just
these very characteristics of a unit. But this way to solve
the problem is of no use for our purposes, because our aim
is not to eliminate a rare derivative, but to minimize the
quantity of tokens, lacking grammatical characteristics, in
the RNC.
5
Naturally, the very frequent cases of the wrong usage of one
correct wordform instead of another correct form, for example,
the wrong usage of the InstrSing of o-nouns instead of GenPlur
and vice-versa, which takes place in Russian texts very often
(e.g., voronom instead of voronov), may be corrected only in
manual mode.
2.1.10. How We solve the Problems
To solve the problems, addressed in the §§2.1.8-2.1.9, we
may act in two ways. First, we may try to expand the
23
The other case of the kind is the usage of the correct, but
rather rare variant of an ending or of a morpheme. For
example, the standard InstrSingFem of the nouns of ideclension is -ju (yarost’-ju < yarost’ ‘fury’). But in the
old and poetic texts it is very easy to find variants with the
ending -iju (yarost’-iju). And this duality is regular. The
RNC parser now leaves the wordforms like yarostiju
without any annotation, that is it treats them like nonstandard. But it’s a mistake – these wordforms are quite
standard in the specific contexts and in the specific types
of texts.
To surmount the grammar obstacle of the kind we have to
improve the grammar basis of our grammar parser. First,
we ought to list all the similar problem points of Russian
Grammar (for example, in the Russian nominal paradigm
the problem points obviously are GenSingMasc, LocSingMasc, NomPlurMasc, GenPlurMasc, InstrSingFem,
and so on). For all these points we must allow all possible
oscillations of the standards, for example, we must recognize the endings GenPlur -ov and -Ø for the nouns like
yabloko (Neuter Inanimate with the velar as the last consonant of a stem) as equal in rights. Naturally, in this
transformed grammar parser the endings -ju and -iju for
the nouns like yarost’ would have equal value.
RNC Vocabulary as much as possible, including all new
derivates and doublets in it. This way is simple, but inefficient: any new text, included in the RNC, may add a derivative or a doublet, which has been absent in the Corpus
until now. The openness is the main feature of productive
derivative sets.
Therefore, the only way to solve the problem is to provide
a special module as a component of the RNC grammar
parser, which would generate the necessary units automatically, regardless of their real presence in the RNC.
For example, every Russian noun with the suffix -enije,
originating to a verb (gorenije ‘burning’ < goret’ ‘to
burn’, burenuje ‘drilling’ < burit’ ‘to drill’, and so on, the
set is open), has a doublet with the suffix -en’je (goren’je,
buren’j). In most cases their meaning are absolutely equal,
let alone the slight stylistic differences, and they have two
regular differences in grammar: in LocSing the nouns
with -enije have the ending -i and the nouns with -en’je
have -e, and in GenPlur the noun with -en’je change their
stems and receive the vowel -i- (not goren’j, but gorenij).
Naturally, to treat these units, there is no need to include
in the RNC Vocabulary all possible pairs – it’s enough to
include one of them and to generate the doublet automatically.
The acute problem of the kind is the open set of composites with the same first or second part of a word, for example, videoplejer ‘video cassette player’, videoinzhener
‘video engineer’, videopirat ‘video pirate’ and so on. It’s
quite obvious that it is absolutely impossible to list all the
derivates of the kind, because the association of two parts
in this case is absolutely free and agglutinative, and the
first part of such composites may combine with all nouns
or/and adjectives. Therefore, in such case we ought to
create the module in the RNC grammar parser, which can
1) divide the composites of the kind into two parts (one of
them – non-variable, the other – variable one), 2) check,
whether the variable part of the composites is present in
the RNC Vocabulary (that is, whether it functions as a
self-reliant unit), and if so, then 3) consider the composite
of this structure as a word-combination, not a word. In
this case, there is no need to bother, whether this or that
composite is present in the RNC Vocabulary – it’s enough
to make sure that the RNC Vocabulary contains the first
and the second parts separately.
2.2.
2.3.
Grammar-NSU
This is the rarest case, when we face the omission in the
grammar of the Corpus: the grammatical phenomenon has
already been described in Russian scientific grammars,
but the creators of the grammar parser (MyStem) have
evaluated this phenomenon as too specific, and so have
been mistaken. We mean the so called new Vocative in
Russian. The forms like mam! (< mama ‘mummy’), pap!
(< papa ‘daddy’), Kol’! (< Kol’a ‘pet name for Nikolay’),
Mash! (< Masha ‘pet name for Mary’), and so on are extremely popular in the informal Russian. Just now we can
find the Vocative only in the Disambiguated Sub-Corpus
of the RNC, where the forms of the kind had been disambiguated in manual mode. In the rest of the RNC the vocative forms are grammaticalized incorrectly.
Therefore, to solve the problem we ought to transform the
grammar parser in the following way. For all the Russian
anthroponyms of the a-declension (i.e. for all names, having the ending -a in NomSing) and for some appellatives
of the same grammatical structure the parser must suppose
one more grammar case, the Vocative. If a name doesn’t
include any suffix, containing an unstable vowel, then the
Vocative of the name coincides with the GenPlur (Masha
Voc=MashaGenPlur=Mash). If a name includes a suffix, containing the unstable vowels -o-/-e-, then the Vocative of
the name coincides with the NomSing, but the last vowel
-a is cut off (MashkaVoc= MashkaNomSing – -a = Mashk6).
Wordform-NSU
This type of non-standard Russian takes place, when the
RNC Vocabulary includes the right lemma and the right
set of necessary grammatical characteristics, but one-toone correspondence between the lemma and the set of its
grammatical features collapses because of the grammatical peculiarity of a certain wordform.
This case takes place when any substandard wordform is
used in a text. For example, in standard Russian the only
normative GenPlur of the lemma yablok-o ‘apple’ is
yablok-Ø. This wordform is fixed in all dictionaries of
Russian. Therefore, when the grammar parser of the RNC
meets in a text GenPlur yablok-ov, it leaves this bastard
form with no annotation. The same situation takes place
when the vowel-variant of the passive/middle verbal postfix -s’a is used after a vowel, in place of the correct consonant-variant -s’, for example, mol’us’a ‘I’m praying’
instead of standard mol’us’. So, in these cases the wordform adopts the alien ending/morpheme.
2.4.
NSU in Spoken and Accentological Corpora
The very specific types of NSU are presented in the two
very important sub-corpora of the RNC – the Spoken
Corpus7 and the Accentological Corpus. These corpora
include the linguistic objects which usually are not reflected in the written Russian. They are the stress and a lot
6
The GenPlur of the names of such structure would include the
unstable vowel: Mashek.
7
About the Spoken Corpus of the RNC see (Grishina, 2006).
24
of peculiarities of oral Russian speech8. It’s widely
known that an ordinary written Russian text isn’t accentuated and it doesn’t contain the linguistic units, which are
highly frequent in the spoken Russian. If that’s the case,
then the RNC grammar parser, which has been expected
to deal only with the written Russian, would be mistaken
dealing with the spoken or accentuated Russian.
If so, then we ought to upgrade the parser and to teach it
to take into account the peculiarities of the spoken Russian. These are the examples.
1) In Russian there are a lot of grammatical forms which
are distinguished only with the help of stress. For example, polúchite ‘you will receive’ vs. poluchíte ‘Receive
it!’, kátera ‘boat, GenSing’ vs. katerá ‘boat, NomPlur’,
Lídok ‘Lidka, informal for the name Lidia, GenPlur’ vs.
Lidók ‘one of the pet names for Lidia, NomSing’, ó ‘1) the
name of the letter O, 2) one of the variant of the particle
vot, 3) the interjection o!’ vs. o ‘the preposition about’,
and so on.
In all these cases the grammar parser must distinguish the
forms and lexemes with/without stress and also the different positions of the stress in a word.
2) The Spoken Russian texts contains the multifarious set
of the contracted (simplified) and apocopic forms of the
words. For example, the very frequent in spoken Russian
words like seychas ‘now’, tys’acha ‘thousand’ have the
standard simplified oral variants schas, scha, tyscha.
Naturally, this kind of pronunciation is reflected in the
scripts, because it’s very important to distinguish, when a
speaker uses the full pronunciation and when he let himself use the simplified forms9.
On the other hand, each unchangeable word in Russian,
which has the vowel ending, may lose this vowel without
any alteration of the meaning of the word. These forms
without the ending vowel are called apocopic and are very
frequent in the spoken Russian (some of them have even
become literary variants, for example, chtoby /chtob ‘in
order to’10.). The scripts of the spoken Russian contain the
great amount of the apocopic forms.
Finally it is necessary to mention the reflection of the dialect peculiarity, which takes place in the scripts of the
spoken Russian, for example, the reflection of the Caucasian accent (sl’ushay instead of slushay ‘listen’, and so
on). These phenomena are extremely frequent in the Dialectal Corpus of the RNC.
All these simplified, apocopic and dialectal variants are
now misadjusted: some of them are skipped over with the
grammar parser and don’t receive any grammar characteristics, some of them receive wrong grammatical qualities
(for example, the apocopic variant of the particle pr’am –
instead of pr’amo ‘real, really, exactly’ – is analyzed as
the NomSing of the short form of the adjective pr’amoj
‘straight’).
To solve the problems, concerned in this section, we must
act in different ways. When a phenomenon is rare and
sporadic, or when it takes place only in the highly specific
types of texts (for example, in the dialectal texts), the only
way to grammaticalize these bastard forms is the preserving annotation, started in manual mode (see §2.1.7). When
the phenomena are frequent and regular, we ought to provide their presence in the RNC Vocabulary. For example,
the frequent apocopic and simplified form are to be included in the Vocabulary. Similarly, the regular contracted forms of the Russian masculine patronymic ought
to be included in the Vocabulary as the variants of the full
forms: the oral variant Ivanych < Ivanovich, Nikolaich <
Nikolajevich ‘the son of Ivan or Nikolay’ can be obtained
in the course of the regular transformations of the initial
patronymics (-ov- → Ø, -i- → -y- OR -jev- → Ø).
3. Conclusion
Thus, we may summarize the previous description and
review all the types of the normalization of the NSU in the
RNC.
1. Expanding of the RNC Vocabulary, sources:
• Existing dictionaries [Proper names (2.1.1); Loanwords (2.1.3); Obsolete and rare words (2.1.4);
Substandard vocabulary (2.1.5); Simplified and
apocopic form (2.4)]
• 1.2. RNC itself [Abbreviations (2.1.2)]
• 1.3. Regular derivation [Russian patronymics
(standard form) (2.1.6)]
2. Improvement of the RNC grammar parser
• 2.1. Creating of the special analyzing module
[Word-forming doublets (2.1.8, 2.1.10); Open sets
of derivatives (2.1.9, 2.1.10); Russian patronymics
(informal forms) (2.4)]
• 2.2. Reforming of the grammar parser [Wordform-NSU (2.2); Grammar-NSU (2.3)]
3. Preserving annotation
• In manual mode [Orthographical corruptions
(2.1.7); Graphical corruptions (2.1.7); Spoken and
accentological NSU (2.4)]
• Orthoparser [Orthographical corruptions (2.1.7)]
Acknowledgements
The work of the RNC group is supported by the program
“Philology and Informatics” of Russian Academy of Sciences. The author’s investigation is supported by the
RFBR (RFFI) under the grants 06-06-80133a and 08-0600371а.
References
Chae Young-Soog (1998). Improvement of Korean Proofreading System Using Corpus and Collocation Rules /
Language, Information and Computation (PACLIC12),
18-20 Feb. P. 328-333
Crowdy, S. (1994). Spoken Corpus Transcription // Literary and Linguistic Computing. 9(1). P. 25-28;
Elistratov, V.S. (2000). Slovar’ russkogo argo. M.
Gramm (supplement) (2003) – Zalizn’ak, A.A. Grammaticheskij slovar’ russkogo jazyka. Prilozhenie: Imena
sobstvennye. M.
Gramm (2003) – Zalizn’ak, A.A. Grammaticheskij slovar’
russkogo jazyka. M.
Grishina, E. (2006). Spoken Russian in the Russian National Corpus (RNC) // LREC2006: 5th International
Conference on Language Resources and Evaluation.
ELRA, 2006. P. 121-124
8
About the peculiarities of a spoken speech and their reflection
in the written texts see (Panunzi, 2004).
9
The creators of the BNC were the first to list the basic difficulties of the orthographical record of a spoken speech and the main
types of the discrepancies between ‘written’ written speech and
‘spoken’ written speech (see, for example, (Crowdy, 1994)).
10
About the usage of the apocopic forms in the spoken Russian
see (Grishina, 2007).
25
Kompjuternaya Lingvistika i Intellektual’nyye Tekhnologii. Trudy Mezhdunarodnoj Konferencii “Dialog 2007”.
Bekasovo, Rossiya. P. 422-427.
NKRJA (2005). Natsional’nyj korpus russkogo yazyka:
2003–2005. Rezul’taty i perspektivy. M.
NKRJA (2008). Natsional’nyj korpus russkogo yazyka:
2006–2008. Rezul’taty i perspektivy. (forthcoming)
Ortho (2005). Russkij orfograficheskij slovar’. M.
Panunzi, A., and al. Using PiTagger for Lemmatization
and PoS Tagging of a Spontaneous Speech Corpus: COral-Rom Italian // LREC2004: 4th International Conference On Language Resources And Evaluation. May
2004. Lisbon, Portugal
Petrovsky, N.A. (1996). Slovar’ russkih lichnyh imen. M.
SIS (2008). Grishina, E.A. Novejshij illustrirovannyj slovar’ inostrannyh slov. M. (forthcoming)
Grishina, E. (2007). Text Navigators in Spoken Russian. //
Proceedings of the workshop “Representation of Semantic Structure of Spoken Speech” (CAEPIA’2007,
Spain, 2007, 12-16.11.07, Salamanca), Salamanca.
P. 39-51.
Nadeau, D. (2007). Semi-Supervised Named Entity Recognition: Learning to Recognize 100 Entity Types with
Little Supervision. Thesis submitted to the Faculty of
Graduate and Postdoctoral Studies in partial fulfillment
of the requirements for the PhD degree in Computer
Science Ottawa-Carleton Institute for Computer Science School of Information Technology and Engineering University of Ottawa. Ottawa.
Nevzorova, O., Nevzorov V., Zin’kina, Ju., P’atkin, N. Integral’naya Tekhnologiya Razresheniya Omonimii v
Sisteme Analiza Tekstovykh Dokumentov LOTA //
26
Creating a Parallel Treebank of the Old Indo-European Bible Translations1
Dag T. T. Haug, Marius L. Jøhndal
University of Oslo
[email protected], [email protected]
Abstract
In this paper, we report on the creation of a syntactic annotation scheme to be used in a comparative study of the oldest extant versions of
the New Testament in Indo-European languages: the Greek original, as well as the translations into Latin, Gothic, Armenian and Church
Slavonic. The tag set was created in a pilot study involving syntactic annotation of the Gospel of Mark in Greek and Latin. The resulting
tag set is well suited for capturing syntactic variation between these languages, particularly in areas having to do with pragmatics and
information structure – as the treebank is created within a larger project in this field – but also more general syntactic differences.
1. Introduction
scheme must therefore be suitable for representing the general structure of sentences in these languages.
It was decided that rather than focussing on creating coherent data from the very start of the annotation process,
the best way to accomplish our objectives was to annotate a pilot text while we were developing the annotation
scheme. This way we could maximize the value of feedback from annotators, gain experience with the annotation
process itself, and have a readily available testbed during
development of the software. The remainder of this paper
describes this process and its outcome.
The project Pragmatic Resources of Old Indo-European
Languages (PROIEL) aims to study the linguistic means of
information structuring which are offered by the grammar
of Greek, Latin, Armenian, Gothic and Church Slavonic,
i.e. the means that the lexicon and the syntax of these languages make available2 for expressing such categories as
old and new information, contrast, parallelism, topicality
and others. Five particular phenomena will be examined in
the PROIEL project:
• Word order
• The definite article
• Discourse particles
• Anaphoric expressions, including zero anaphora
• Participles and absolutive constructions
These topics were chosen because they are known to be important in information structure systems cross-linguistically
and because they are areas where the languages in the corpus are likely to diverge. For example, Ancient Greek is
the only language in the corpus to have a grammaticalized
definite article. This language is also well known for its
abundance of discourse particles, which cannot be rendered
directly in the target languages. Word order is notoriously
free in these languages, and while this led to direct adoption of the Greek word order in many cases, there are still
patterns that cannot be rendered directly. Similarly, the anaphoric and participial systems vary widely.
The most important objective for our treebank is to be able
to represent these phenomena correctly with as fine-grained
information as possible. On the other hand, it is likely that
in the course of the project, we will find other phenomena
that are relevant to the general topic of information structure, so we need to be prepared to adapt our scheme to changing requirements. Finally, it is important that the treebank
is created in such a way as to be useful for a wider audience,
no matter what topics they are interested in. The annotation
2. Preparing the pilot text and creating the
annotation tools
Building a treebank is labour-intensive, so our initial concerns were to avoid duplication of efforts and to get our annotators started as quickly as possible. This was greatly facilitated by the availability of a morphologically annotated
electronic version of the Greek New Testament (SandborgPetersen, 2008) and by the work done by the Perseus digital
library (Crane, 1987; Crane et al., 2001) on their electronic
version of Jerome’s Vulgate and word-lists for Latin and
Greek.
We used these resources to prepare the text for the pilot
study. This text consists of the Greek and Latin versions of
the Gospel of Mark – which in each language amounts to
roughly 13,000 words or 10% of the complete New Testament.
Due to the complexity of Biblical textual criticism, and
since the purpose of the overarching project is to do a crosslinguistic comparative study, we chose to ignore manuscript
variants. Our texts are instead based on the text of a specific edition, and we only correct digitization errors, should
these occur.
The preparation and annotation of the pilot text proceeded
in four stages:
• Pre-processing
• Automated morphological tagging
• Manual annotation by annotators
• Manual review by a reviewer
The pre-processing stage involved segmentation, detection
of sentence boundaries and sentence alignment. Segmentation is occasionally problematic as certain morphemes behave as separate entities in the syntactic model we use, but
1
The research project described here is funded by the Norwegian Research Council’s YFF program, grant no. 180632. We
thankfully acknowledge this support. The glossing in this paper
follows the Leipzig Glossing Rules (http://www.eva.mpg.
de/lingua/resources/glossing-rules.php).
2
Prosody presumably also played a role which is, however,
irrevocably lost for us.
27
Major part of speech
Verb
form part of other words. This happens, for example, with
instances of krasis3 in Greek. A trickier case is presented in
Latin where enclitics like -que cannot always be tokenized
unambiguously.
For detecting sentence boundaries, we decided to use punctuation as a guide, as the canonical division into verses frequently fails to match what we consider to be meaningful
syntactic units. A particular problem was presented in our
Vulgate text as it lacks punctuation altogether. To solve
this, the punctuation from a different electronically available edition, the Clementine Vulgate, was merged into our
text by first aligning the orthography of the two editions,
then using an implementation of the standard diff algorithm
(Hunt and McIlroy, 1976) and finally patching the target
text using only chunks that involved insertion of punctuation. This simple approach gave good results in spite of
numerous textual differences, and only occasionally resulted in off-by-one errors in sentence division.
To answer the questions posed in the research project, corpus users will eventually need to be able to query the same
sentence in multiple languages. This requires sentences
to be aligned, and our experiments indicate that Gale and
Church (1993)’s algorithm performs this task well with
chapters as hard delimiters and sentence boundaries as soft
delimiters.
As ready-made word-form lists were available, morphological tagging was done simply as an assignment of one or
more pairs of lemma and morphological tags to each token
in the text. The morphological tag set that we developed
is more fine-grained than those of our source data, so for
many tokens the level of detail of the assigned morphological tags was insufficient. We were able to address this
by manually adding assignment rules, and harvesting additional rules from data already disambiguated by annotators.
The morphological tags are positional tags and derived
from the system used by the Latin Dependency Treebank
(Bamman and Crane, 2006). For the purposes of the
PROIEL project, more fine-grained distinctions had to be
made for certain parts of speech, in particular pronouns,
whose functions are primarily pragmatic. We have also deviated from the traditional grammarian’s view of morphology and adopted a more ‘modern’ view by treating adverbs
that double as prepositions as intransitive prepositions, and
by merging the two categories particle and adverb (see table
1).
It has furthermore been necessary to introduce a means for
indicating ambiguity of form or unresolvable syncretism,
e.g. for nouns that alternate between genders. This has
been solved by allowing multiple inheritance within each
field of a positional tag, so that, for example, the tag for
masculine gender has a super-tag that indicates masculine
or neuter gender, one that indicates masculine or feminine
gender etc.
The two manual stages of the process – annotation and review – were performed using variants of the same graphical
interface. We wanted an interface that students could use on
Noun
Pronoun
Numeral
Minor part of speech
declinable common noun
indeclinable common noun
declinable proper noun
indeclinable proper noun
relative pronoun
interrogative pronoun
indefinite pronoun
demonstrative pronoun
personal pronoun
possessive pronoun
personal reflexive pronoun
possessive reflexive pronoun
reciprocal pronoun
declinable cardinal number
indeclinable cardinal number
ordinal number
Adjective
Article
Adverb
comparable adverb
relative adverb
interrogative adverb
other non-comparable adverb
Conjunction
Subjunction
Preposition
Interjection
Foreign word
Table 1: The parts of speech defined in the PROIEL morphological tag set.
typical campus workstations which frequently have a limited choice of installed software and restrict students’ rights
to run local, stand-alone applications. We therefore chose
to develop a light-weight web-based interface that would
function with only a modern browser and client-side scripting.
The interface is designed as an incremental ‘wizard’ that
splits the annotation into three steps. First, annotators
verify and, if necessary, adjust sentence boundaries. We
have found that this ought to be restricted as annotators felt
tempted to override the judgements of the text editors and
therefore did excessive adjustments of sentence boundaries. This may be due to the style of our texts in which
sentences tend to ‘run together’ and a large number of sentences are introduced by and. The choice between coordinating a main clause with the preceding clause or not is thus
often an arbitrary one. We therefore let annotators adjust
the sentence boundaries only one token at a time so that
they could only correct off-by-one errors.
The second step of the ‘wizard’ involves morphological
disambiguation. Annotators are presented with the output
from the automated morphological tagging and are asked
to choose the correct lemma and tag pair in cases of ambiguity. Finally, in the third step, annotators build depend-
3
The term krasis refers to a contraction in which the final
vowel of one word coalesces with the initial vowel of the next,
and the two words are written together.
28
ency structures for each sentence. This is done visually and
guided by a simple rule-based ‘guesser’ that suggests the
most likely dependency relation based on the morphology
of head and dependent.
In addition to the interface for annotation, and an interface for text browsing, we added functionality for tracking
change history and for inserting cross-references to other
information sources such as dictionaries. In particular, we
have made use of Strong’s Concordance (Strong, 1890) and
the Analytical Lexicon of the Greek New Testament (Friberg
et al., 2000), as these were the basis for lemmatization in
our Greek text.
The system is based on Ruby on Rails with a database backend. As a by-product of this choice, the system offers not
only a traditional web-interface to the corpus, but also exposes a RESTful XML interface that can be used by clients
to query the database. This should facilitate interchange of
data and direct reuse of our work in other contexts.
pragmatics interface, because they are both optional and
receivers of a thematic role from the verb. 4
In the adnominal domain, it is well known that there are
interesting correlations between types of genitives and information structure. For example, possessive genitives tend
to be old information in a text and are typically used to access new referents, whereas object genitives are more often
new information. Partitive genitives are special as they, and
not their syntactic heads, introduce the discourse referent of
a noun phrase: ‘two of the disciples’ refer to a group of disciples, and not to some kind of ‘twoness’, unlike ‘the teaching of the disciples’. It is therefore essential for PROIEL to
distinguish these uses of the genitive.
There is one notable exception to the general pattern that
our tags are more fine-grained than those of the LDT; the
LDT scheme provides 9 subtypes of auxiliary relations:
AuxP for prepositions, AuxC for conjunctions, AuxR for
the reflexive passive etc. In our opinion all items with the
relation AuxX in the LDT can be conflated to a single relation as instances can still be differentiated based on lexical
information when the need arises.
3. The development of the annotation
scheme
As noted above, all the languages in our corpus have a
‘free’ word order, i.e. the word order does not indicate syntactic dependencies or grammatical functions, but serves
pragmatic purposes. Therefore, while word order data are
important for PROIEL, they cannot be conflated with information about grammatical function as is done in a phrase
structure grammar. For this reason, it was decided to base
the annotation scheme on dependency grammar (DG). This
also had the advantage that other projects developing treebanks of Latin, e.g. the Latin Dependency Treebank (LDT),
are based on DG, using a faithful adaption of the welldocumented Prague Dependency Treebank (PDT) (Hajič,
1998).
We began our work using the Greek and Latin versions of
the New Testament, since these exist publicly available in
electronic form with morphological annotation. We expected the syntax of most old Indo-European languages to be
sufficiently similar to be captured within a single annotation
scheme and our experienced with the Greek and Latin texts
have confirmed this. There are diverging constructions, of
course, but they can all be captured using our primitive syntactic relations, and we do not expect Gothic, Armenian or
Church Slavonic to be different in this respect.
3.2. Granularity
By asking annotators to do fine-grained classification of the
data, we run the risk of more inconsistencies in the application of the scheme. For this reason, we have introduced
some ‘super-tags’, i.e. tags that we ask the annotators to
use whenever they are in doubt. For example, it can be hard
to tell whether a given relative clause is restrictive (ATR) or
not (APOS). We provide a tag REL for such cases, so that
the annotators do not simply guess.
However, in the case of adnominal tags, we purposefully
did not provide any such super-tag, in order to test the viability of making distinctions within this domain. The results
were mixed. In the beginning, we asked annotators simply
to distinguish attributes and appositions. After a couple
of weeks, we introduced more granularity by means of the
tags PART, to be used for partitive expressions, and OBL,
to be used whenever an expression is an argument of the
noun – typically an object genitive as in amor fati ‘love of
faith’. OBL was chosen because this is the relation we use
for non-object arguments in the verbal domain.
When the pilot was finished, we studied how annotators had
used these tags. Although the results are not statistically
significant, they were valuable in guiding our development
of the annotation scheme. In general, the annotators coped
well with the PART relation: of 42 uses of this relation, only
3 were wrong – not too bad a result at such an early stage in
the annotators’ training. Moreover, the errors could easily
be detected automatically, since they did not involve any
uses of PART with an adnominal genitive that should have
had another relation, but rather the generalisation of PART
to other contexts with partitive semantics, i.e. a genitive
object and an object of the Greek preposition apo.
The concept of arguments of nouns was harder to apply.
This relation was used 22 times for items dependent on
3.1. General presentation
While we wanted to keep the option to automatically convert our treebank to a more general format, we soon realised that the level of granularity of the PDT annotation
scheme or the LDT annotation scheme (Bamman et al.,
2007) would not be sufficient for PROIEL. Table 2 shows
the general outline of our annotation scheme in comparison
with that used by LDT. It is more fine-grained than the LDT
scheme, both in the domain of verbal arguments and that of
adnominal functions. To study the interaction between syntax/argument structure and pragmatics in determining word
order, we need to be able to separate objects (OBJ) from
other arguments of the verb (OBL). Furthermore, agent expressions (AG) are particularly interesting for the syntax-
4
The decision to include the AG relation, which combines syntactic function and semantic role, was a pragmatic choice motivated by the fact that we do not expect to have the resources to do
a full tectogrammatical annotation as in the PDT.
29
Latin Dependency Treebank
PRED
*
SBJ
OBJ
ADV
ATR
ATV
PNOM
OCOMP
COORD
APOS
AuxX (X defines the subtype of Aux)
ExD
PROIEL Corpus
PRED
PRED
SUB
OBJ
OBL
AG
XOBJ
ADV
ATR
NARG
PART
XADV
XOBJ
XOBJ
*
APOS
Aux
*
VOC
Explanation
Main clause predicate
Subordinate clause predicate
Subject
Object
Oblique
Agent
Open complement clause
Adverbial
Attribute
Nominal argument
Partitive
Free predicative
Subject complement
Object complement
Coordinator
Apposition
Auxiliary
External dependency
Vocative
Table 2: Sentential functions in LDT and PROIEL. An asterisk in one of the columns indicates that the two annotation
schemes diverge in some other way than by one simply being more specific than the other.
the adnominal domain, we have introduced a super-tag ADNOM so that we no longer force the annotators to choose
when they are in doubt.
nouns, 6 times erroneously. Apparently the possibility of
using OBL adnominally tempted annotators into analysing
verbs with an object and a PP complement as if the prepositional phrase were dependent on the object, e.g. so that
super is a dependent of manus in the participial construction
3.3. Dealing with covert elements
A dependency grammar is well equipped to deal with the
free word order of ancient languages. However, it is less
well adept at representing another feature typical of old
Indo-European languages, namely ellipsis. The DG formalism has difficulties with all constructions without a clear
syntactic head, e.g. ellipsis, coordination (and in particular
asyndetic coordination) and sentences lacking a verb (most
often the copula).
Different solutions have been devised to these problems;
in the following we describe our solution, which tries to
capture the facts in a theory-neutral manner.
First consider a less well-known problem for dependency
grammar, namely ellipsed dependents. Ellipsis of dependents is much more frequent than ellipsis of heads and, while
it is generally easier to deal with, it can sometimes lead to
problems. Consider the treatment of the accusative with infinitive (AcI) (example 2) and the complement infinitive in
the LDT (example 3):
(1) imponens
manus
super
put.PRS . PTCP. NOM . SG hand.ACC . PL upon
illos
they.ACC . PL .
‘laying his hands upon them’
Such errors cannot be detected automatically. Moreover,
since the OBL tag is used in more contexts, we run the additional risk of contaminating the entire set of OBL-relations.
Not only was OBL used in cases where another relation
should have been used, there were also cases where PART
and OBL were not used when they should have been. As
part of our analysis of the data from the pilot annotation, we
examined the 123 cases of genitive nouns dependent on another noun that had been annotated after the introduction of
PART and OBL as adnominal tags. 17 of these were given
an incorrect analysis, and in 16 cases this was because ATR
was used when PART or OBL would have been correct.
The period of pilot annotation has taught us that it is difficult for annotators to distinguish different functions in the
adnominal domain. Still we will continue to make these
distinctions, but we no longer use the relation OBL, but
rather a separate relation NARG (nominal argument) which
is devoted to arguments of nouns. In this way, we have
an ‘exit strategy’ in case the the quality of the annotation
remains low, since we can merely convert all NARGs to
ATRs. Also, since we have now had the opportunity to test
the annotators’ ability to make fine-grained distinctions in
(2)
Root
PRED
dixit
SUB
puer
OBJ
vivere
SUB
matrem
30
puer
dixit
matrem
boy.NOM . SG say.3 SG . PFV mother.ACC . SG .
vivere
live.PRS . INF
‘The boy said his mother was alive.’
(3)
verbs in full sentences, which would be sisters of their arguments, and verbs in participial constructions, which would
dominate their arguments. Although this system provided
an intuitive way of dealing with so-called ‘gapping’ (the absence of the verb in the second conjunct, see example 12),
we quickly abandoned it.
Inspired by Lexical-functional grammar, we instead chose
to represent the structural difference between infinitives in
AcIs and complement infinitives as two contrasting relations, OBJ and XOBJ. The latter function is by definition
one which cannot have an overt subject, but shares its subject with another element in the clause. We designate this
structure-sharing by what we call ‘slash notation’.7 The full
representation of example 3 is therefore:
Root
PRED
voluit
SUB
OBJ
mater
proficisci
(5)
mater
voluit
proficisci
mother.NOM . SG . can.3. SG . PFV travel.PRS . INF
‘Mother wanted to travel.’
Root
PRED
voluit
The fact that we here have two different constructions is
signalled only by the presence of a subject daughter in example 2. However, Latin being a pro-drop language,5 this
subject is optional:6
(4)
SUB
XOBJ
mater
Root
proficisci
mater
voluit
proficisci
mother.NOM . SG . can.3. SG . PFV travel.PRS . INF
‘Mother wanted to travel.’
PRED
putaverunt
The arrow in this example should be interpreted as a secondary dependency relation. In this case it shows that mater
is the subject of both proficisci and voluit. This accounts for
case agreement with predicate nominals in the dependent
infinitive construction, as in the following example, which
also shows how we deal with instances where the subject of
the XOBJ is not overtly realized:
OBJ
esse
PIV
phantasma
(6)
putaverunt
esse
phantasma
think.3. PL . PFV be.PRS . INF ghost.ACC . SG
‘They thought [he] was a ghost.’
Root
PRED
dicitur
A search for all AcI structures would have to sacrifice precision (by matching all infinitive objects) or recall rate (by
matching all infinitives that have a subject daughter). This
problem is particularly important to PROIEL, since the subject of the infinitive in such examples as example 4 can only
be left out because it is given information in the context.
At first we tried to solve the problem by not letting the verb
stand in for the whole sentence, but rather let sentences (including AcIs) be represented by an empty node that dominated the verb and its arguments, so that the defining feature
of these empty nodes was the possibility (but not necessity)
of dominating a subject. However, this quickly leads to
problems: the empty elements are hard to deal with computationally and result in an unmotivated distinction between
SUB
Marcus
XOBJ
esse
XOBJ
aeger
Marcus
dicitur
aeger
Marcus.NOM . SG say.3. SG . PRS . PASS ill.NOM . SG .
esse
esse.PRS . INF.
‘Marcus is said to be ill.’
In traditional grammar the subject of aeger is supplied by
the verb esse. We incorporate this by letting the slash arrow
point to the head verb whenever it ‘ought’ to pointed to a
‘pro-dropped’ argument. This has the further advantage of
5
The term pro-drop language refers to languages in which
some pronouns may be omitted when they can be inferred pragmatically.
6
In this tree, we use PIV for the traditional notion of ‘predicative complement’, which actually does not exist in our scheme;
see below.
7
This designation is in turn inspired by the vaguely similar
Head-driven phrase structure grammar.
SLASH -lists of
31
making it easier to validate annotations; we can enforce the
principle that every XOBJ or XADV relation should have
one slash arrow and that this arrow should point towards the
head verb or an element dominated by the verb.
esse in turn gets its subject from Marcus. Notice that we
treat the traditional category of predicative complement as
XOBJ, seeing that the facts are the same: the element is
subcategorized for by the verb and does not have a direct
relation to its subject.
Our representation is neutral between control and raising
analyses. Compare the example above to the following example:
(7)
ille
respondens
he.NOM . SG answer.PRS . PTC . NOM . SG
dixit
eis
say.3. SG . PFV they.DAT. PL
‘Answering them, he said’
(9)
Root
PRED
dixit
OBL XADV
Root
eis
PRED
respondens
dixit
answer.PRS . PTC . NOM . SG say.3. SG . PFV
eis
they.DAT. PL
exarkesei
OBL XOBJ
Note that this preserves the structure even when the subject
of the predicative participle is ‘pro-dropped’. This is otherwise hard to achieve: If the participle were to depend on its
subject, the structure would be different when the subject
was omitted. We could have used complex tags to denote
the subject of the participle, but then we would have had to
deal with predicative participles that depend on main clause
adjuncts, of which there may be several. The slash notation
solves this by pointing to the subject of the participle and
not to the relation which the subject of the participle has.
Dependency grammars generally speaking only allow coordination of elements related to the same head via the same
relation. It is, however, not hard to find examples with a
conjunct participle coordinated with an adverbial element,
e.g. an absolute genitive:
genesthai
soi
XOBJ
turannon
exarkesei
soi
genesthai
suffice.3. SG . FUT you.DAT become.PFV. INF
turannon
tyrant.ACC . SG
‘It will suffice for you to become a tyrant.’
Many theories would treat example 6 as raising and example 7 as control, to explain the differences in case agreement. Our representations are meant only as input to such
discussions, so we make no commitment regarding the
status of our slash dependencies, nor do we ask our annotators to make decisions based on linguistic theory. The
annotators should simply identify the subject of infinitives
and predicative complements: if the subject of the XOBJ is
not present in the subtree dominated by the governing verb,
the slash arrow should point to the verb, which ‘supplies
the subject’ according to traditional grammar.
(10)
Root
PRED
eisin
XOBJ
3.4. Other uses of the slash notation
The slash notation was first introduced to separate AcI’s
from complement infinitives even in cases where the subject of the AcI has been ‘pro-dropped’. It is a modest but
powerful addition to our data-model that allows us to obtain
a richer annotation for many structures. The slash notation
most manifestly translates to infinite predications that are
not arguments of the main verb, i.e. predicative (conjunct)
participles:
(8)
respondens
ADV
asthenees
kai
XADV
ADV
helkomenoi
epileipontôn
AG
SUB OBJ
Root
upo
ombron
OBL
AUX
hêliou
tôn
PRED
dixit
SUB OBL XADV
ille
eis
respondens
AUX
32
tou
autous
(12)
tôn
ombrôn
the.GEN . PL rain.GEN . PL .
epileipontôn
autous
kai
leave behind.PRS . PTCP. GEN . PL they.ACC . PL and
upo tou
hêliou
by the.GEN . SG sun.GEN . SG
helkomenoi
asthenees
draw up.PRS . PTCP. NOM . PL weak.NOM . PL
eisi
be.3. PL . PRS
‘With the rain leaving them being and drawn up by
the sun, they [sc. the rivers] are weak.’
Root
PRED
et
PRED
frater
SUB
et
SUB
SUB
OBJ
qui
aurem
SUB OBJ
fratrem
in
pater
filium
The combination of a restricted use of empty nodes and the
slash notation makes it possible to preserve the structure
of the tree. We also capture the fact that the argument in
mortem is shared between the two conjuncts. The two slash
arrows have rather different interpretation: the one from the
empty node to the verb indicates sharing of lexical material,
whereas the one from the empty node to the preposition in
indicates a double dependency.
Since the slash relation is not labelled, it is important that
the relation can be interpreted based on other information
in the sentence. And, in fact, this remains possible. We can
distinguish three uses of the slash notation.
• Slash arrows from an empty node to a sister node signal predicate identity
• Slash arrows from an XOBJ or XADV node to a
mother or sister node indicate the subject of the infinite
verb
• Slash arrows from other verbal nodes signal a shared
argument
The first case is not a dependency relation at all, so there is
no need to infer a label. In the second case, the slash arrow
always indicates a SUB relation: there is widespread typological support for ‘controlled’ functions always being subjects, and this holds for the old Indo-European languages
as well. Only the third kind of slash arrow may have different labels. We therefore constrain such arrows to cases
where the shared arguments have the same function in both
conjuncts. This is by far the most frequent case. The following example illustrates how unambiguous interpretation
is possible even in complex cases:
audiat
habet
SUB OBJ OBL
tradet
frater
fratrem
deliver.3. SG . PRS brother.NOM . SG brother.ACC . SG
in mortem
et pater
to death.ACC . SG and father.NOM . SG
filium
son.ACC . SG
‘The brother shall betray the brother to death, and the
father the son.’
PRED
vivit
0
mortem
Root
SUB
tradet
OBL
Our notation solves this problem. The daughter nodes of
XADV relations always have a slash arrow, so the ‘X’
merely serves to indicate the presence of the slash.8 This
means that we can coordinate XADV and ADV without distorting the analysis. The advantage of our notation would
become even clearer if the participle had an overt subject, as
the two conjuncts in this case would have to have different
heads in a ‘classical’ analysis.
Once introduced, the slash notation can be exploited for
richer annotation of other structures that involve ellipsis or
structure-sharing. Since the verb substitutes for the whole
sentence, we treat coordination of two verbs as sentence
coordination and use the slashes to indicate double dependencies (i.e. subject sharing):
(11)
PRED
qui
vivit
et aurem
who.NOM . SG live.3. SG . PRS and ear.ACC . SG
habet
audiat
have.3. SG . PRS hear.3. SG . PRS . SBJV
‘Whoever lives and has ears shall hear.’
The advantage of this notation is evident in gapping constructions where the predicate is omitted in the second conjunct:
8
The relation between XOBJ and OBJ is of another nature
since verbs subcategorize differently for OBJ and XOBJ.
33
(13)
Root
PRED
kai
PRED PRED
edeiran
apesteilan
XADV
XADV
labontes
kenon
OBJ
auton
kai labontes
auton
edeiran
and take.PRS . PTCP he.ACC . SG beat.3. PL . PFV
kai apesteilan
kenon
and send away.3. PL . PFV emtpy.ACC . SG .
‘Having captured him, they beat him and sent him
away empty-handed.’
In this graph, we capture the information that auton is an
object of all verbs in the sentence; that the subjects of the
free predicatives labontes and kenon are elided arguments
of the verbs edeiran and apesteilan; and that labontes is an
adverbial adjunct (here, in fact, equivalent to a subordinate
temporal clause) relevant to both main verbs.
Thus the simple addition of an extra binary relation in our
data model enables us to capture a wide variety of facts
about structure sharing without introducing a plethora of
empty nodes. Notice also that our two levels of annotation
are not interdependent: while the slash arrows cannot be
interpreted without the dependency tree, the opposite does
not hold. If in some processes (such as automated parsing)
we are forced to exclude slash relations, the dependency
tree can still be drawn independently and the slash relation
added by other means.
The annotators made two kinds of errors in dealing with
the slash arrows: sometimes they forgot to use them where
they should have been used, and sometimes they attached
arrows indicating double dependencies to an empty node.
In example 12, they introduced an empty OBL-node under
the empty verb in second conjunct, and a slash arrow from
the OBL-node to the preposition in. In this way, they enforced a more consistent interpretation of the slash notation
as an indication of identity of lexical material. We considered this option, but rejected it due to the proliferation
of empty nodes it leads to.
Fortunately, both these errors are easily detectable, and as
future extension the annotation interface will enforce validation constraints that prohibit dependency graphs that have
an XOBJ or XADV nodes lacking a slash arrow, or in which
a slash arrow exits an empty node which has been assigned
a non-PRED relation.
4. Conclusion
While work on the complete PROIEL corpus is still in its 34
infancy, we feel that the pilot stage of the project has en-
abled us to establish a relatively firm base for the annotation
scheme to be used and the accompanying tools that annotators will rely on. In the course of our work, we have been
confronted with many of the ‘classical’ difficulties that syntactic theory still struggles with, e.g. the difficulty of strict
morphological categorisation and the analysis of ellipsis in
Dependency Grammar, but also encountered novel problems that arise in ancient Indo-European languages. Our
solutions to these problems should enable us to address the
needs of the PROIEL project, but still be sufficiently theory
neutral to ensure that the corpus will be useful for others,
and the technologies used should enable an open exchange
of data and eliminate many obstacles for potential reuse of
our data.
5.
References
David Bamman and Gregory Crane. 2006. The design
and use of a Latin dependency treebank. In Jan Hajič
and Joakim Nivre, editors, Proceedings of the Fifth International Treebanks and Linguistic Theories, pages
67–78, Prague. Data available from http://nlp.
perseus.tufts.edu/syntax/treebank.
David Bamman, Marco Passarotti, Gregory Crane, and
Savina Raynaud. 2007. Guidelines for the syntactic annotation of Latin treebanks. Technical report, Tufts Digital Library, Medford. Version 1.3.
Gregory Crane, Robert F. Chavez, Anne Mahoney,
Thomas L. Milbank, Jeffrey A. Rydberg-Cox, David A.
Smith, and Clifford E. Wulfman. 2001. Drudgery and
deep thought: Designing digital libraries for the humanities. Communications of the ACM, 44(5):34–40.
Gregory Crane. 1987. From the old to the new: Integrating hypertext into traditional scholarship. In Hypertext
’87: Proceedings of the 1st ACM conference on Hypertext, pages 51–56. ACM Press.
Timothy Friberg, Barbara Friberg, and Neva F. Miller.
2000. Analytical Lexicon of the Greek New Testament.
Baker, Grand Rapids.
William A. Gale and Kenneth W. Church. 1993. A program for aligning sentences in bilingual corpora. Computational Linguistics, 19(1):75–102.
Jan Hajič. 1998. Building a syntactically annotated corpus: The Prague dependency treebank. In Eva Hajičová,
editor, Issues of Valency and Meaning. Studies in Honor
of Jarmila Panevová, pages 12–19. Charles University
Press, Prague.
James W. Hunt and M. Douglas McIlroy. 1976. An algorithm for differential file comparison. Computing Science Technical Report 41, Bell Laboratories.
Ulrik Sandborg-Petersen. 2008. Tischendorf’s 8th edition Greek New Testament with morphological tags.
Version 2.0. http://morphgnt.org/projects/
tischendorf.
James Strong. 1890. The exhaustive Concordance of the
Bible : showing every word of the text of the common English version of the Canonical Books. Methodist
Book Concern, New York.
Integrating Language Technology in a web-enabled Cultural Heritage system
Penny Labropoulou1, Harris Papageorgiou1, Byron Georgantopoulos1, Dimitra Tsagogeorga1,
Iason Demiros2, Vassilios Antonopoulos2
1
Institute for Language and Speech Processing
Epidavrou & Artemidos, Athens, Greece
2
Qualia Technologies of Understanding
Voriou Ipirou 4, Athens, Greece
{ penny, xaris, byron, dtsag } @ilsp.gr ; { idemiros, vantonop } @qualia.gr
Abstract
This paper describes a web-enabled sophisticated Cultural Heritage (CH) system giving access to digital resources of various media,
which exploits Language Technologies (LT) in order to enhance the performance of the search and retrieval mechanisms. More
specifically, the paper presents the system requirements and architecture, drawing aspects from: (a) the cultural data repository and its
particularities; (b) the unified metadata scheme that has been devised, integrating elements from various metadata standards, providing
thus a rich description of the resources; (c) the thesauri (one of the major pillars of the system) that provide uniform access to the
resources. The LT that form part of the system construction and use are presented in detail, focusing on the Term Extraction and Named
Entity Recognition tools used in the construction of the thesauri and the metadata annotation process, and the Term Matching module
exploited in the mining process for the identification of query terms which appear in a morphosyntactically similar form in the thesauri.
•
1.
and, finally, a Video Collection (VC), including
audiovisual material depicting Asia Minor
settlements now and then.
As evidenced by the above description, the material not
simply pertains to different media but, in fact, constitutes
five distinct archives, which have been collected and
documented at different times by different people to a
varying degree of granularity and according to different
classification schemes, although they all address the same
subject. The main goals of the project were to:
• digitize the selected material to a format suitable for
long-term preservation;
• document the data with any information deemed
necessary for its preservation and exploitation,
catering for the interrelation of the resources so as to
highlight their common features and allow unified
access to the whole set;
• make the resources publicly available to all interested
parties, ranging from the research community to
laypersons such as Asia Minor descendants, school
students and people interested in finding out more
about the particular historical period.
Introduction
In this paper we present an integrated system currently
under development which aims to provide web-enabled
access to digitized Cultural Heritage (CH) resources of
various media. The paper describes the main components
of the system and the way these are exploited to facilitate
users’ access. Section 2 presents the cultural resources
handled in the project and the metadata scheme that has
been implemented to describe them. Section 3 presents
the system architecture and Section 4 the thesauri that are
integrated in the system as a separate module. Finally,
Section 5 discusses the Language Technology (LT)
methods and tools that have been (a) exploited in the
course of the project to facilitate the system construction
and (b) integrated in the search and retrieval mechanisms
to improve the system performance. In the last section, we
present conclusions and future work.
2.
Cultural resources collection &
metadata standards
2.1 Description of the resources
2.2 The Metadata Scheme
The CH system presented in this paper aims at providing
web-enabled sophisticated access to the digital cultural
resources of the Centre for Asia Minor Studies (CAMS).
The institute's resources range from paper manuscripts to
3-D objects related to the everyday life cycle of Greeks in
Asia Minor from the end of the Byzantine era until the
1922 divide. From these, only a subset has been selected
for digitization:
• the Archive of Oral Accounts (AOA), which
includes "testimonies" of refugees recorded on paper
by field researchers in the 1930's - 1940's;
• the Manuscript Archive (MA), made up of
manuscripts containing refugees' reminiscences;
• the Photographic Archive (PA), with photos
picturing everyday life in Asia Minor, the 1922
events and the resettlement in Greece;
• the Map and Sketches Archive (MSA), with maps
and cartographic sketches drawn by the Cartographic
Service;
The core feature of the platform we have developed for
the access and exchange of the CAMS cultural resources
lies on the metadata scheme used for their documentation.
The CAMS metadata scheme is composite, adopting as a
base
the
Encoding
Archival
Description
(EAD, http://www.loc.gov/ead/) enriched with elements
drawn from various metadata standards for digital
resources (e.g. Dublin Core (http://dublincore.org/) for
the whole set, TEI (http://www.tei-c.org/index.xml) for
the MA, DIG35 (http://www.i3a.org/i_dig35.html) for the
PA, etc.), adapted to the project's needs and supplemented,
where required, with the institute's particular
requirements. In this way, the CAMS scheme succeeds in
encoding different types of information regarding the
form and content of the original resources and their digital
substitutes, in a uniform representation, regardless of the
medium format of each particular item.
The adopted methodology follows considerations
reported in (Duval et al, 2002) stating: "In a modular
35
metadata world, data elements from different schemas as
well as vocabularies and other building blocks can be
combined in a syntactically and semantically
interoperable way". The resulting scheme has been
implemented as an Application Profile and incorporated
in the Metadata Annotator Workbench ("MetaWork")
which is integrated within the overall Service Oriented
Architecture of our platform. The MetaWork interface
provides annotators with a powerful tool for efficient
metadata collaborative editing of digitized cultural
resources and search thereof.
3.
4.
Thesauri description
The use of thesauri in knowledge mining systems has well
proven its merits (Carpineto et al., 2001). In the CH
domain, thesauri can be exploited to mine knowledge
hidden in textual data (combined with LT tools) or, in the
simplest scenario, to improve access to the resources.
In CH systems, thesauri are commonly used as part of the
annotation metadata scheme providing classification
information. The two CAMS thesauri are also used for the
same purpose but, in contrast to the usual practice, they
are kept outside the annotation scheme. This has been a
deliberate choice: an important requisite of the project
was the ability to open up to the national and international
research community, by providing the appropriate
infrastructure for establishing links with other parties
collecting similar material; by keeping the thesauri
content independent of the annotation scheme, they can
be easily merged/replaced/supplemented by other thesauri
and term lists or linked with meta-thesauri, allowing for
seamless cross-collection access, without jeopardizing the
annotation scheme. Linking with the appropriate
documents is performed at the annotation scheme level.
The thesauri are integrated in the query system providing
access to the resources with which the terms are linked.
As such, they are of utmost importance to the system
performance. Until now, users interested in CAMS
resources had access to the material through the
intermediary of specialized personnel, who have full
knowledge of the contents and form of the material and
who could process the users' queries so as to present them
with the most appropriate material in response to their
queries. This process is replaced in the web-based
services by the thesauri: the query system "interprets" the
users' queries through the thesauri in order to mine the
best plausible results in the digital repository.
Access to the resources via the thesauri is performed in
the form of free text input in the query system, or by
browsing through the thesaurus viewer, which presents
terms hierarchically or alphabetically sorted. The user can
choose between viewing the detailed term card or the
associated digital resources.
System architecture
The web-enabled CH system that provides access to the
CAMS material draws elements from the following pools
of resources:
• the digital substitutes of the CAMS originals, that
have been authorized for web access;
• the annotation metadata accompanying them;
• two thesauri implementing the classification schemes
of the cultural material.
A Service-oriented architecture connecting various
components and subsystems that are loosely coupled and
easily managed has been implemented (Figure 1). The
architecture enables the seamless integration of different
components that are situated behind the portal and
represented as services. At the front end, users interact
with a single composite service that aggregates LT and
back-end data sources. User friendliness, on the one hand,
and utmost functionality and informativeness, on the other,
have been the guiding principles behind its development.
The user interfaces allow querying on a specific subset of
the metadata which has been carefully selected by the
CAMS personnel to reflect the needs of intended users.
Querying on this subset is performed by free text input
and value selection from lists, depending on the query
element. The two most important query elements are
related to the classification of the material, which falls
along two axes, namely geographical and thematic,
implemented in the form of thesauri.
Cultural Resources Collection
Security And Authentication Metadata Schema Thesauri Interfaces Interoperability & exchange Oral Archives
Manuscripts
•testimonies
•...
•autobiographies
•...
Map Archive
Photo Archive
•photographs
•...
Sketches
•maps
•...
•cartographic sketches
•...
Video Collection
•audiovisual material
•...
Digitization & Protection process
Metadata Editor Workbench [MetaWork]
Storage Metadata Database
Indexing Module Retrieval Engine
Query Processing Thesauri & Term matching Services
User Interfaces / Portal
Figure 1: System architecture
36
In both modes, the thesauri relations 1 are exploited to
improve the system performance: in the browsing mode,
the user can navigate from higher to lower nodes, from
non-preferred to preferred terms and/or to related terms;
in the search mode, the system is responsible for query
expansion utilizing the relations in order to improve recall
and precision, a methodology well established in the
bibliography of information retrieval systems.
The thematic thesaurus has been constructed in such a
way that it could be expanded in the future so as to include
other material of the institute itself or an external source.
In this endeavour, the Ethnographic Thesaurus
(http://et.afsnet.org/about.html) has served as our
"prototype", i.e. a subset of the ET facets have been
selected for our purposes, and a similar structure for the
hierarchical organization of the terms has been adopted.
The thesaurus contains approximately 1,100 terms in each
of the three project languages (Greek, English, and
Turkish).
The geographical thesaurus includes ca. 4,500 terms, of
which about 2,500 represent preferred terms. This is due
to the existence of multiple alternative names used for the
same place, reflecting the coexistence and interaction of
the Turkish and Greek populations (and others, of course)
in the area; this variety is further accentuated by the
presence of various dialects and idioms rendering more
linguistic and/or spelling variants. The organization of
this thesaurus has followed the original system devised by
CAMS, because it reflects the situation in the time period
under investigation: some of the settlements included in
the material do not exist any longer, while others have
changed names; therefore, it is extremely difficult to
attribute the names with absolute certainty to current
locations.
The development of both thesauri has been assisted by a
term extraction tool described in Section 5.1.
5.
Thus, a backup plan that we decided to explore was to
apply the TE procedure to an electronic corpus of relevant
texts 2 . The corpus consisted of texts provided by CAMS,
consisting of descriptions of the resources (e.g.
promotional material, newspaper articles) and research
work based on them (mainly academic articles from the
CAMS Bulletin and scientific journals), as well as texts
from internet sources which, however, addressed only a
small part of the CAMS themes (mainly the Asia Minor
expedition and tragedy); a very small subset is taken
directly from the annotation metadata, namely the
questionnaire used in the construction of the AOA. Finally,
the idea of performing multilingual TE was quickly
abandoned because the vast majority of the texts were in
Greek, very few in English and none in Turkish.
The thesaurus construction procedure followed in the
course of the project consisted of the following steps:
(a) pre-processing of the corpus, i.e. tokenization,
below-part-of-speech (POS) tagging and lemmatization;
(b) Named Entity (NE) recognition, catering for the
identification of named locations (the "terms" included in
the geographical thesaurus);
(c) TE aiming at identifying candidate terms for both
thesauri;
(d) manual validation of the list of candidate terms by
subject experts, aiming at selection of valid terms;
(e) processing of the term list in order to add terms and
establish the appropriate relations between them; for the
purposes of the project, this step included the task of
adding the translation equivalents.
Analysis
ILSP Term Recognition
Tokenization
Tagging/Lemma
LT deployment in the system
Text
Existing LT tools and techniques have been deployed in
the framework of the project to improve access to the
cultural material in two ways:
(a) during the construction process, to minimize the time
and cost effort of building certain system modules,
namely the thesauri and the annotation metadata and
(b) in the form of components integrated in the system,
to enhance its performance: two modules, a stemmer
and a term matching module are activated in the
query subsystem.
Pattern Grammar
rules
Finite Automata
conversion
Named Entity
Recognition
Parsing with
Pattern
Grammar
Statistical
Filtering
Candidate
terms
Figure 2: Term extraction workflow
Figure 2 shows the automatic processing stages used in
the above procedure. The LT tools employed are:
• the ILSP tokenizer, that performs tokenization and
identification of sentence boundaries and other
surface linguistic phenomena (punctuation, dates,
numbers etc.);
• the ILSP tagger (Papageorgiou et al., 2000), which
consists of a transformation-based POS tagger
(compatible with the relevant EAGLES standards)
and a lexicon-based lemmatizer: the tagger assigns
initial tags to known words by looking up in a lexicon
created from the manually annotated corpus during
training; then, a suffix-lexicon is used for initially
tagging unknown words; a set of contextual rules are
5.1. LT in the system development
The construction of a thesaurus is an important time- and
effort-cost procedure. To minimize this cost, Term
Extraction (TE) methods are recommended for term
harvesting from digitized resources. In CH systems, such
resources are the textual data contained in the written
documents and the metadata of audiovisual items.
However, in this project, this could not be the case
because the written documents are not digitized at text
level while term-intensive annotation metadata (e.g. titles
of manuscripts, captions of photographs etc.) were not
available at the time of the thesauri construction.
2
Relevant texts should not be overlooked in TE since they
contain alternative modes of expressing similar notions which
qualify as "non-preferred terms" for access to the material.
1
The form of the thesauri is conformant to the ISO2788 and
ISO5964 standards for monolingual and multilingual thesauri.
37
Terms in thesauri are usually kept in one single form,
called canonical. However, they can be encountered in
various other forms which may differ from the canonical
form in morphological and other aspects. In addition,
spelling mistakes, word omissions, changes in functional
words, in word order etc. will not allow a term to be
located in its exact form in a predefined term list.
Matching problems arise when a system lacking linguistic
processing capabilities is used for looking up terms in
various forms against a thesaurus. This derives from the
fact that crude string comparison techniques will fail in
matching terms that are morphologically or in other ways
varied. A more sophisticated technique is needed in order
to identify the closest canonical term or terms (ideally the
term) that the user originally was looking for. This process
is called term matching.
As aforementioned, the problem of alternative names is
especially accentuated in the geographical thesaurus; for
instance, nowadays Izmir is also encountered under the
names of Smyrna, Smirna, Smyrni, Smirni etc., related to
naming at different time periods and variance in the
transcription systems into English. In addition, we should
stress the fact that the names of the settlements (a large
number of which are of Turkish origin) have been
recorded by the institute's field researchers, who often had
no knowledge of Turkish, and who tried to put into the
Greek alphabet what they understood from the refugees'
pronunciation. Thus, the original Turkish name may have
been paraphrased into Greek (either because the Greek
inhabitants of Asia Minor used a slightly modified version
of the original name or simply because the recording was
not the most appropriate one). In this situation, term
matching is of great value to the query enhancement,
since it allows the system to locate closely similar forms
(paraphrases) of the same word. In the thematic thesaurus,
its use is less evident but still appreciated, since it allows
for finding spelling variants without entering them in the
thesaurus.
The term matching algorithm takes as input a query term
and returns a ranked set of the closest canonical terms. It
operates in two pipelined stages:
1. Word matching: The words of the query term are
matched against the wordforms occurring in the thesaurus.
An error tolerant finite-state recognizer is used to
recognize all the strings in the regular set plus any strings
that are produced from the regular set by a small number
of unit editing operations. The edit distance between two
strings measures the minimum number of unit editing
operations of insertion, deletion, substitution and
transposition of adjacent symbols required to convert the
one string into the other (Oflazer, 1996). The algorithm,
being string-oriented, is generic enough to be applicable
to all the three languages of the thesaurus. For example,
ed(Φουντουκλού, Φουντουκλιά)=2 and ed(Smyrna,
Smirni)=2.
then applied to improve the initial phase output;
finally, the lemmas are retrieved from a Greek
morphological lexicon containing 70K lemmas;
• the MENER NE recognition system (Giouli et al.,
2006): it is a highly modified version of the
best-scoring system that participated in the
CoNLL-2003 shared task developed by Hai Leong
Chieu and Hwee Tou Ng. MENER adopts a
single-level maximum entropy approach that
combines sentence-based local evidence about words,
with global evidence, which is a collection of features
drawn from other occurrences of those words within
the same document;
• a TE module developed at our institute
(Georgantopoulos & Piperidis, 2000): it is a hybrid
system comprising:
o corpus parsing of below-POS tagged texts, based
on a term pattern grammar endowed with regular
expressions and feature-structure unification,
which produces single- and multi-word candidate
terms;
o statistical filtering, using the TFIDF metric, in
order to remove grammar-extracted terms
lacking statistical evidence; TFIDF is a standard
weight computation method which favours terms
having high frequency in the current document
and low frequency in a reference corpus
(collection of documents); the reference corpus is
a general language news corpus of about ~10K
words in size for Greek.
All the resources are semi-automatically linked to the
appropriate term in the metadata annotation process:
some of the metadata (among which, titles of manuscripts
and videos as well as captions of the PA) have been made
available at later stages of the project in digital form in
simple document files 3 ; these have been lemmatized and
checked for inclusion of terms from the thesauri; the links
established in this procedure will be checked by the
annotator during the final annotation process.
In addition, the NE module has been used for the
identification of named persons in the captions of
photographs; these will be coded in the relevant
photographs' metadata, providing links to the metadata
associated with other resources: e.g. a person that appears
on a photograph may be an interviewee in the AOA or the
writer of a manuscript in the MA.
5.2. LT in the query system
The query system benefits from two modules that enhance
its performance: a stemmer and a term matching module.
Stemming is the technique that has been selected for the
free text queries (in Greek and English); the reasons for
having selected stemming over morphological lexica is
that it can be easier implemented for other languages
(especially for English), while the metadata vocabulary,
although rather limited, contains mainly domain specific
words which are not included in general language lexica.
The most important LT module integrated in the system is
the Term Matching module. This is used to enhance the
thesauri query subsystem.
3
Had these metadata been available earlier, their use in the TE
procedure would have greatly contributed to the results, since
they are term-intensive data for the particular subject.
38
In order to avoid the over-generation of useless
suggestions having too many spelling mistakes, we have
imposed a maximum allowable number of errors t, where
t’s value is analogous to the word length. After
experimentation, we have found the following values for t
that demonstrate best results:
length
1-4
5-7
8–9
10+
with the query term by sorting them according to the word
matching distances.
Figure 3 illustrates the term matching process when the
input query is "Ano xorio". The best two terms contain
word matches, though not exact, for both query words,
"Ano" and "xorio".
6.
t
1
2
3
4
The matching procedure examines every possible path
from the start node to the final nodes, in depth-first mode,
so that the parsed strings deviate no more than t editings
from the string to be matched. The algorithm is very
efficient since the edit distance of each new (longer)
string is recursively calculated from the edit distances of
its substrings.
2. Term matching: We recall for each word retrieved in
the previous step the set of terms where it is encountered
in the thesaurus. Combinatorial processing is performed
on the set of candidate terms in order to decide which
thesaurus terms match best the query term, based on a
penalty-assigning mechanism illustrated in the following
formula:
7.
∑
References
Carpineto, C., De Mori, R., Romano, G., Bigi, B. (2001).
An information theoretic approach to automatic query
expansion. ACM Transactions on Information Systems,
Vol 19, n. 1, pp. 1--27.
Duval, E., Hodgins, W., Sutton, S. & Weibel, S.L. (2002).
Metadata principles and practicalities. D-Lib Magazine,
8(4). Available: doi:10.1045/april2002-weibel.
Georgantopoulos, B. & Piperidis, S. (2000). A Hybrid
Technique for Automatic Term Extraction.
Proceedings of the ACIDCA 2000 Conference.
Giouli, V., Konstadinidis, A., Desipri, E., Papageorgiou,
H., Piperidis, S. (2006). Multi-domain Multi-lingual
Named Entity Recognition: Re-visiting & grounding
the resources issue. International Conference on
Language Resources & Evaluation, LREC 2006.
Oflazer, K. (1996). Error-tolerant Finite State Recognition
with Applications to Morphological Analysis and
Spelling Correction. Computational Linguistics
Volume 22 No.1.
Papageorgiou, H., Prokopidis, P., Giouli, V., Piperidis, S.
(2000). A Unified POS Tagging Architecture and its
application to Greek. International Conference on
Language Resources & Evaluation, LREC 2000,
pp.1455--1463.
n
term- penaltyi =
Conclusions and future work
In this paper, we have described a CH system giving
access to heterogeneous digitized resources. We have
focused on the benefits of exploitation and integration of
LT both at the metadata annotation process and the search
and retrieval interface as well as the challenges such an
endeavour imposes. In the near future, we intend to
evaluate the system in quantitative and qualitative terms.
(edit-penaltyij)
j =1
where n is the number of words in the query term, term
penalty is the penalty of the i-th candidate term,
edit-penalty is the pre-computed edit distance of the j-th
word in the i-th term (or a high constant penalty if no
match exists). Finally, the terms are ranked according to
their penalty, in ascending order.
The scoring mechanism effectively computes the join of
all the canonical terms that contain at least one of the
query words (verbatim or variants). It then classifies these
terms through the penalising schema: exact word matches
do not get a penalty score, error tolerant matches are
penalised with their respective edit distance and, finally,
terms are heavily penalised for every query word they do
not contain. As a result, good scoring terms are the ones
that contain word matches for all or most of the query
words. Terms containing all of the query words are
guaranteed to outrank all the other candidates. We further
filter terms having an equal number of common words
39
Morphological Tagging of Old Norse Texts
and Its Use in Studying Syntactic Variation and Change
Eiríkur Rögnvaldsson1, Sigrún Helgadóttir2
Department of Icelandic, University of Iceland1, Árni Magnússon Institute for Icelandic Studies2
{Árnagarði við Suðurgötu, IS-1011, Neshaga 16, IS-1072}, Reykjavík, Iceland
E-mail: {eirikur,sigruhel}@hi.is
Abstract
We describe experiments with morphosyntactic tagging of Old Norse narrative texts using different tagging models for the TnT tagger
(Brants, 2000) and a tagset of almost 700 tags. It is shown that by using a model that has been trained on both Modern Icelandic texts
and Old Norse texts, we can get 92.7% tagging accuracy which is considerably better than the 90.4% that have been reported for
Modern Icelandic. In the second half of the paper, we show that the richness of our tagset enables us to use the morphosyntactic tags in
searching for certain syntactic constructions and features in a large corpus of Old Norse narrative texts. We demonstrate this by searching for – and finding – previously undiscovered examples of two syntactic constructions in the corpus. We conclude that in an inflectional language like Old Norse, a morphologically tagged corpus like this can be an important tool in studying syntactic variation and
change.
1.
Introduction
2.
In a previous project (Helgadóttir 2005; 2007), we have
trained the TnT tagger written by Brants (cf. Brants, 2000)
on a corpus of Modern Icelandic. The corpus used in that
project was created in the making of the Icelandic Frequency Dictionary (Íslensk orðtíðnibók, henceforth IFD;
Pind et al., 1991). The IFD corpus is considered to be a
carefully balanced corpus consisting of 590,297 tokens
with 59,358 types – both figures including punctuation.
The corpus contains 100 fragments of texts, approximately 5,000 tokens each. All the texts were published for
the first time in 1980–1989. Five categories of texts were
considered, i.e. Icelandic fiction, translated fiction,
biographies and memoirs, non-fiction (evenly divided
between science and humanities) and books for children
and youngsters (original Icelandic and translations).
The texts were pre-tagged using a specially designed
computer program and the tagging was then carefully
checked and corrected manually. Thus, this corpus is ideal
as training material for data-driven statistical taggers,
such as the TnT tagger.
In the present project, we applied the TnT tagger trained
on the Modern Icelandic corpus to Old Norse (Old Icelandic) texts. 1 This paper describes the results of this
experiment, and also describes our experiments with
using the morphologically tagged Old Norse corpus to
search for syntactic constructions. We conclude that in an
inflectional language like Old Norse, a morphologically
tagged corpus like this can be an important tool in
studying syntactic variation and change.
Tagging Modern Icelandic
In this section, we describe the tagset used in our research,
and give a brief overview of our experience with the training of the TnT tagger on Modern Icelandic texts.
2.1 The tagset
The tagset developed for the IFD corpus is very large,
compared to tagsets designed for English at least, such as
the Penn Treebank tagset (Marcus et al., 1993). The size
of the tagset of course reflects the inflectional character of
Icelandic, since it is for the most part based on the traditional Icelandic analysis of the parts of speech and grammatical categories, with some exceptions where that
classification has been rationalized.
In the tag strings, each character corresponds to a single
morphosyntactic category. The first character always
marks the part of speech. Thus, the sentence Hún hefur
mætt gamla manninum ‘She has met the old man’ will be
tagged like this:
(1) Hún
hefur
mætt
gamla
manninum
fpven
sfg3eþ
ssg
lkeþvf
nkeþg
The meaning of the tags is as follows:
(2) fpven: pronoun (f) – personal (p) – feminine (v) –
singular (e) – nominative (n)
sfg3eþ: verb (s) – indicative (f) – active (g) – 3rd person (3) – singular (e) – past (þ)
ssg:
verb (s) – supine (s) – active (g)
lkeþvf: adjective (l) – masculine (k) – singular (e) –
dative (þ) – definite (v) – positive (f)
nkeþg: noun (n) – masculine (k) – singular (e) – dative (þ) – suffixed article (g)
1
It is customary to use the term ‘Old Norse’ for the language
spoken in Norway, Iceland and the Faroe Islands up to the
middle of the 14th century. The overwhelming majority of
existing texts written in this language is either of Icelandic origin
or only preserved in Icelandic manuscripts. For the purposes of
this paper, ‘Old Norse’ is thus synonymous with ‘Old Icelandic’.
40
Of the word forms in the IFD corpus, 15.9% are ambiguous as to the tagset within the IFD. This figure is quite
high, at least compared to English, which reflects the fact
that the inflectional morphology of Icelandic is
considerably more complex than English. Icelandic nouns
can have up to 16 grammatical forms or tags, verbs up to
106 different tags, and adjectives up to 120 tags. Altogether, 639 different tags occur in the IFD corpus, but the
total sum of possible tags is around 700.
Some of the ambiguity is due to the fact that inflectional
endings in Icelandic have many roles, the same ending
often appearing in many places (e.g. -a in penna for all
oblique cases in the singular (acc., dat., gen.), and
accusative and genitive in the plural of the masculine
noun penni ‘pen’, producing 5 different tags for one form
of the same word). The most ambiguous of word forms in
the IFD, minni, has 24 tags in the corpus, and has not exhausted its possibilities (Bjarnadóttir, 2002).2
the experiment on Swedish text (Megyesi, 2002) and
indicates that the major difficulty in annotating Icelandic
words stems from the difficulty in finding the correct tag
for unknown words. Words belonging to the open word
classes (nouns, adjectives and verbs) account for about
96% of unknown words in the test sets whereas words in
these word classes account for just over 51% of all words
in the test sets.
3.
Having trained the TnT tagger on Modern Icelandic texts,
we wanted to find out whether the tagger could be of help
in tagging Old Norse narrative texts, with the purpose of
facilitating the use of these texts in research on syntactic
variation and change. To create a manually annotated
training corpus for Old Norse from scratch would have
been a very time-consuming task. Thus, the possibility of
using the bootstrapping method that we describe in this
section was a key factor in realizing this project.
Bootstrapping is of course a common approach in training
taggers and parsers. To our knowledge, however, this
approach has not been used in historical linguistics to
develop tagging models for a different stage of language
than the tagger was originally trained on. Our method
somewhat resembles the experiments of Hwa et al. (2005),
who used parallel texts to build training corpora by
projecting syntactic relations from English to languages
for which no parsed corpora were available. The training
corpora created using this method were then in turn used
to develop stochastic parsers for the languages in question.
The whole process took only a small fragment of the time
it would have taken to create a manually corrected corpus
to train the parsers.
The common factor in our project and the work reported
by Hwa et al. (2005) is the use of another language, or (in
our case) another stage of the same language, as a starting
point in the bootstrapping process. Our experiments with
bootstrapping the tagging of Old Norse texts are described
in this section.
2.2 Training the tagger
The computer files for the IFD corpus each contain one
text excerpt. Each file was divided into ten approximately
equal parts. From these, ten different disjoint pairs of files
were created. In each pair there is a training set containing
about 90% of the tokens from the corpus and a test set
containing about 10% of the tokens from the corpus. Each
set should therefore contain a representative sample from
all genres in the corpus. The test sets are independent of
each other whereas the training sets overlap and share
about 80% of the examples. All words in the texts except
proper nouns start with a lower case letter.
Results for ten-fold cross-validation testing for the TnT
tagger are shown in table 1 (cf. Helgadóttir, 2005; 2007).
It is worth noticing that these results show lower
performance rates when the tagger is applied to the Icelandic corpus than is achieved for example for Swedish as
reported in Megyesi (2002). In that study, TnT was applied to and tested on the SUC corpus with 139 tags
compared to the Icelandic tagset of almost 700 tags.
Performance rates are also considerably lower than have
been reported for the systems trained on the Penn treebank.
Type
All words
Known words
Unknown words
Tagging Old Norse texts
3.1 Old Norse vs. Modern Icelandic
At a first glance, it may seem unlikely that a tagger trained
on 20th century language could be applied to 600-700
years old texts. However, Icelandic is often claimed to
have undergone relatively small changes from the oldest
written sources up to the present. The sound system,
especially the vowel system, has changed dramatically,
but these changes have not led to radical reduction or
simplification of the system and hence they have not
affected the inflectional system, which has not changed in
any relevant respects. Thus, the tag set developed for
Modern Icelandic can be applied to Old Norse without
any modifications.
The vocabulary has also been rather stable. Of course, a
great number of new words (loanwords, derived words
and compounds) have entered the language, but the
majority of the Old Norse vocabulary is still in use in
Modern Icelandic, even though many words are confined
to more formal styles and may have an archaic flavor.
Accuracy %
90.4
91.7
71.6
Table 1: Mean tagging accuracy for all words, known
words and unknown words for TnT.
Table 1 shows results for known words, unknown words
and all words. Mean percentage of unknown words in the
ten test sets was 6.84. This is similar to what was seen in
2
minni can be a noun meaning ‘memory’, present tense of the
verb minna ‘remind’, comparative of the (irregular) adjective
lítill ‘small’. In all of these words we find extensive syncretism,
resulting in many different tag strings for this word form in each
part of speech.
41
250,920
preserved in vellum manuscripts from the 13th through the
15th centuries, but some of them only exist in paper
manuscripts from the 16th and 17th centuries. This makes
it extremely difficult to assess the validity of these texts as
linguistic evidence, since it is often impossible to know
whether a certain feature of the preserved text stems from
the original or from the scribe of the preserved copy, or
perhaps from the scribe of an intermediate link between
the original and the preserved manuscript. It is well
known that scribes often did not retain the spelling of the
original when they made copies; instead, they used the
spelling that they were used to. In many cases, two or
more manuscripts of the same text are preserved, and
usually they differ to a greater or lesser extent. Furthermore, it is known that not all of the editions that our
electronic texts are based on are sufficiently accurate (cf.,
for instance, Degnbol, 1985).
Even though this may to some extent undermine the validity of the texts as sources of syntactic evidence, it does
not directly concern the main subject of this paper, which
is to show that we can use a tagging model developed for
Modern Icelandic to assist us in making the Old Norse
corpus a usable tool in studies of syntactic variation and
change. There is no reason to believe that possible
inaccuracies and errors in the texts – cases where they fail
to mirror correctly the syntax of the manuscripts – have
any effects on the tagging accuracy. That is, the use of
more accurate editions would not lead to less accurate
tagging.
42,745
3.3 Training the tagger on the Old Norse corpus
On the other hand, many features of the syntax have
changed (cf. Faarlund, 2004; Rögnvaldsson, 2005). These
changes involve for instance word order, especially within
the verb phrase, the use of phonologically “empty” NPs in
subject (and object) position, the introduction of the
expletive það ‘it, there’, the development of new modal
constructions such as vera að ‘be in the process of’ and
vera búinn að ‘have done/ finished’, etc.
In spite of these changes, we found it worthwhile to try to
adapt the tagging model that we had trained for Modern
Icelandic to our Old Norse electronic corpus. Our motive
was not to get a 100% correct tagging of the Old Norse
texts, but rather to facilitate the use of the texts in syntactic research, cf. Section 4 below.
3.2 The Old Norse corpus
Our Old Norse corpus consists of a number of narrative
prose texts (sagas), which are assumed to have been written in the 13th and 14th centuries – a few of them probably
later. Among these are many of the most famous Old
Norse sagas. The division of the corpus is shown in Table
2:
Text
Family Sagas (around 40 sagas)
(Íslendingasögur)
Sturlunga Saga
(“Contemporary Sagas”)
Heimskringla
(Sagas of the Kings of Norway)
The Book of Settlement
(Landnámabók)
Total
Tokens
1,074,731
283,002
We started by running TnT on the whole Old Norse corpus using the tagging model developed for Modern Icelandic (cf. Helgadóttir, 2005; 2007). We then measured
the accuracy by taking four samples of 1,000 words each
from different texts in the corpus – one from the Family
Sagas, one from Heimskringla, and two from Sturlunga
Saga – and checking them manually. Counting the correct
tags in these samples gave 88.0% correct tags, compared
to 90.4% for Modern Icelandic.
Even though these results were worse than those we got
for Modern Icelandic, we considered them surprisingly
good. The syntax of Old Norse differs from Modern Icelandic syntax in many ways, as mentioned above, and one
would especially expect the differences in word order to
greatly affect the performance of a trigram based tagger
like TnT. However, sentences in the Old Norse corpus are
often rather short, which may make them easier to analyze
than the longer sentences of Modern Icelandic.
We then selected seven whole texts (sagas) and two fragments from the Sturlunga collection for manual correction
– around 95,000 words in all. This amounts to one third of
the Sturlunga collection. The manual correction was a
time-consuming task, but the time and effort spent on
checking and correcting the output of TnT was only a
small fragment of the time and effort it would have taken
to tag the raw text.
We trained TnT on the corrected text (95,000 words),
tagged the whole corpus again with the resulting model,
1,651,398
Table 2: Division of the Old Norse corpus.
The texts we use are (with the exception of The Book of
Settlement) taken from editions, which were published
between 1985 and 1991 (Halldórsson et al., 1985-86;
Kristjánsdóttir et al., 1988; Kristjánsdóttir et al., 1991). In
these editions, the text has been normalized to Modern
Icelandic spelling. This involves, for instance, reducing
the number of vowel symbols (‘æ’ is used for both ‘ae
ligature’ (æ) and ‘oe ligature’ (œ), ‘ö’ is used for both ‘o
with a slash’ (ø) and ‘o with a hook’), inserting u between
a consonant and a word-final r (maðr ‘man’ > maður),
shortening word-final ss and rr (íss ‘ice’ > ís, herr ‘army’
> her), changing word-final t and k in unstressed syllables
to ð and g, respectively (þat ‘it’ > það, ok ‘and’ > og), etc.
Furthermore, a few inflectional endings are changed to
Modern Icelandic form.
It must be emphasized, however, that these changes do not
in any way simplify the inflectional system or lead to the
loss of morphological distinctions in the texts. Thus, the
texts are just as good as sources of syntactic evidence as
texts that are published in the normalized Old Norse spelling.
On the other hand, we must point out that the original
versions of these texts do not exist; the texts are mostly
42
and measured the accuracy on the same four samples of
1,000 words each as in the first experiment. Now the results were much better – 91.7% correct tags, which is
better than the 90.4% accuracy that we got for Modern
Icelandic. It may seem surprising how much the accuracy
improved when we used this model, especially when we
consider that the training corpus was much smaller than
the training corpus for Modern Icelandic (95,000 words
compared to more than 500,000). On a closer look, however, this is understandable.
First, many of the errors occurring in the first experiment
could be predicted and were easy to correct. For instance,
the word er was always classified as a verb in the third (or
first) person singular present indicative (‘is, am’), as it
usually is in Modern Icelandic. In Old Norse, however,
this word is very often a temporal conjunction (‘when’) or
a relative particle (‘that, which’). When the tagger was
trained on a corrected Old Norse text, it could quickly and
easily learn the correct tagging of these words, due to their
frequency.
Second, it is well known that tagging accuracy is usually
very much lower for unknown words than for known
words, and the number of unknown words was much
lower in the second experiment. In the first experiment,
using the model for Modern Icelandic, the unknown word
rate was 14.6%, reflecting the fact that a number of Old
Norse words are rare or do not occur in Modern Icelandic.
In the second experiment, using the model for Old Norse,
the unknown word rate dropped to 9.6%, even though the
training corpus was much smaller as pointed out above.
This reflects the relatively small vocabulary of the Old
Norse texts, which in turn reflects the narrow universe
that the texts describe (cf. also Rögnvaldsson, 1990).
Finally, we trained TnT on a union of the corrected Old
Norse texts and the Modern Icelandic texts. Thus, the
training set for the final experiment consists of around
500,000 words from Modern Icelandic texts plus 95,000
words from Old Norse texts. When we tagged the Old
Norse corpus using this model, we got 92.7% accuracy for
the same four samples as in the first two experiments. The
results of the three experiments are shown in Table 3:
Tagging model
Modern Icelandic model
Old Norse model
MI + ON model
a separate tag. Thus, if we assume that the tag is correct
when all three models agree, we only need to look at
15.4% of the whole corpus. This means that the highest
possible accuracy to be obtained using this method is
96.3%, since all models agree on a wrong tag in the remaining cases as pointed out above.
We could also choose to disregard the model that is
trained only on Modern Icelandic texts, since it gives
much lower accuracy than the other two models. The remaining models agree on the tagging of 93.5% of the
words – incorrectly for 4.3% of the words. If we only look
at the 6.5% where the models disagree, we are down to
around 107,000 words that we have to correct manually.
This is a manageable task, which we intend to finish in the
near future. We think that performance may exceed 95%
after manual revision of the training set, assuming that
about half of the disagreements can be correctly resolved.
This is an acceptable result in our view, and should be
sufficient for most uses of the corpus.
In this connection, it must be pointed out that a majority
of the tagging errors only involve one morphosyntactic
feature. Thus, nouns are often tagged as accusative instead of dative, or vice versa, whereas gender and number
are correctly tagged; verbs are often tagged as 3rd person
instead of 1st person, whereas mood, voice, number, and
tense are correctly tagged; etc. This means that by using
fuzzy search, we should in many cases be able to find
what we are looking for, even if the words are not quite
correctly tagged.
4.
Tagged texts in syntactic research
Over the past two decades, interest in historical syntax has
grown substantially among linguists. Accompanied by the
growing amount of electronically available texts, this has
led to the desire for – and possibility of – creating syntactically parsed corpora of historical texts, which could
be used to facilitate search for examples of certain
syntactic features and constructions. A few such corpora
have been developed, the most notable being the Penn
Parsed Corpora of Historical English, developed by Anthony Kroch and his associates (Kroch and Taylor, 2000;
Kroch et al., 2004). These corpora have already proven
their usefulness in a number of studies of older stages of
English (cf., for instance, Kroch et al., 2000; Kroch and
Taylor, 2001).
We wanted to know whether our tagged Old Norse corpus
could be used in syntactic research in a similar manner as
syntactically parsed corpora. We had been using the raw
unannotated texts for this purpose (cf., for instance, Rögnvaldsson, 1995; 1996) but the search for certain syntactic
constructions and features had proven to be cumbersome
and give insufficient results. Although our tagging is
morphological in nature, the tags carry a substantial
amount of syntactic information and the tagging is detailed enough for the syntactic function of words to be
more or less deduced from their morphology and the
adjacent words. Thus, for instance, a noun in the nominative case can reasonably safely be assumed to be a subject,
unless it is preceded by the copula vera ‘to be’ which is in
Accuracy %
88.0
91.7
92.7
Table 3: Tagging accuracy for Old Norse texts using three
different tagging models.
It is possible to improve the results by tagging the texts
using all three models and combining the results of
different models in various ways. All three models agree
on the tags for 84.6% of the words. In 80.9% of the cases,
they agree on the correct tag, but for 3.7% of the words,
all three models agree on a wrong tag.
For 15.4% of the words, the models disagree. In most
cases, two of them assign the same tag and the third model
assigns a different tag. In a few cases, each model assigns
43
turn preceded by another noun in the nominative, in which
case the second noun is a predicative complement. A noun
in the accusative or dative case can in most instances be
assumed to be a (direct or indirect) object, unless it is
immediately preceded by a preposition (cf. also
Rögnvaldsson, 2006). As is well known, Modern Icelandic also has accusative and dative subjects, and even
some nominative objects (Thráinsson, 2007), but these
can easily be identified from their accompanying verbs.
To test the usefulness of the tagging of Old Norse texts in
syntactic research, we have made a small study of two
controversial and disputed features of Old Norse syntax;
Object Shift and Passive. These studies are described in
this section.
DP Object Shift in his study of Old Norse, and neither did
Sundquist (2002) in a study of Middle Norwegian. Thus,
Sundquist concludes that “full DP Object Shift is not an
option in earlier stages of Mainland Scandinavian”.
It is therefore of considerable theoretical interest to search
for examples of full DP Object Shift in Old Norse texts.
However, this is a tedious and time-consuming task. Even
though this is a perfectly grammatical construction in
Modern Icelandic, it appears to be very rare in texts. Thus,
one can read dozens or even hundreds of pages without
finding a single example. When the constructions that we
are looking for are that rare, it is easy to overlook the few
examples that actually occur in the texts that we read.
Given the rarity of full DP Object Shift in Modern Icelandic, one may wonder whether those who have studied
Object Shift in Old Norse have looked at a large enough
corpus.
We have searched for examples of full DP Object Shift in
our morphologically tagged Old Norse corpus. In this
search, we use a simple program that searches for a verb
in the indicative or the subjunctive, followed by a noun,
an adjective, or a demonstrative pronoun in an oblique
case, followed by a negation (one of the words eigi, ei,
ekki ‘not’, aldrei, aldregi ‘never’). We allow for up to two
words between the noun/adjective/demonstrative pronoun
and the negation. Thus, in addition to simple sentences
with a noun immediately following the verb and preceding the negation, we will find sentences where both a
demonstrative pronoun and an adjective precedes the
noun, and sentences where a prepositional phrase consisting of a preposition and a noun follows the head noun.
Of course, we will neither get 100% precision nor 100%
recall by using this pattern. It will miss some potential
examples of Object Shift; for instance, sentences with an
adverb modifying a prenominal adjective when a demonstrative pronoun is also present, or sentences with an
adjective modifying an object of a preposition, which
follows the head noun. Furthermore, this search pattern
will return a number of sentences that are not instances of
Object Shift.
When we run this search pattern on the Old Norse corpus,
it returns 245 examples. The majority of these examples
do not show Object Shift. These are for instance sentences
like (5):
4.1 Object Shift
As originally described by Holmberg (1986), Object Shift
is the process of moving a (direct or indirect) object to the
left across a negation. In Modern Icelandic, this process
applies both to pronouns and full NPs (or DPs), as shown
in (3), whereas in the “Mainland” Scandinavian languages
(Danish, Norwegian, and Swedish), it only applies to
pronouns, as (4) shows (examples from Thráinsson, 2007).
The “shifted” object is underlined whereas the negation is
in boldface and the “place of origin” of the shifted object
is shown by an underscore:
(3) Nemandinn las bókina ekki ___
the student read book not
‘The student didn’t read the book’
Nemandinn las hana ekki ___
the student read she not
‘The student didn’t read it’
(4) *Studenten læste bogen ikke ___
the student read book not
‘The student didn’t read the book’
Studenten læste den ikke ___
the student read she not
‘The student didn’t read it’
It has been suggested that this difference between Icelandic and the Mainland Scandinavian languages is
somehow related to the fact that Icelandic has a much
richer case morphology than the Mainland Scandinavian
languages (cf. Holmberg and Platzack, 1995). If this were
so, one would expect to find both types of Object Shift in
Old Norse, since the case system of Icelandic is in all
relevant respects the same as in Old Norse. The Mainland
Scandinavian languages would then be assumed to have
lost Object Shift of full DPs due to the loss of case
inflections.
However, it has been claimed that Object Shift of full DPs
does not occur in Old Norse. Mason (1999) claims to have
found two examples of shifted full DP objects in his study
of nine Old Norse sagas. Sundquist (2002), on the other
hand, concludes “that these two examples do not provide
evidence for a full DP Object Shift like in modern Icelandic”. Haugan (2001) did not find any examples of full
(5) hann skal þetta fé aldregi fá ___ síðan
‘he shall this money never get since’
‘he shall never have this money again’
In this sentence, the fronted NP þetta fé is not an object of
the verb skal, but rather an object of the verb fá. Thus, this
is not an instance of Object Shift but rather shows OV
order in the VP, which is quite a different matter (see, for
instance, Rögnvaldsson, 1996; Hróarsdóttir, 2000).
However, it doesn’t take long to clean the search results
and throw away the sentences that do not show Object
Shift. When we have finished this cleaning, it appears that
we really are left with some genuine examples of full DP
Object Shift:
44
(6) a. Nú leita þeir um skóginn og finna Gísla eigi ___
now search they about the forest and find Gisli not
‘Now they search through the forest and don’t
find Gisli’
b. er hann dræpi Þórð eigi ___og förunauta hans
when he killed Thord not and companions his
‘if he didn’t kill Thord and his companions’
c. og fundu Þórð eigi ___ sem von var að
and found Thord not as expectance was at
‘and not surprisingly, they didn’t find Thord’
in the literature on this subject. Three of these sentences
are shown below – the agentive phrases in boldface:
(7) a. að Þorvarður Spak-Böðvarsson hafi skírður verið
af Friðreki biskupi
that Thorvard Spak-Bodvarsson has baptized been
by Fridrek bishop
‘that Thorvard Spak-Bodvarsson has been baptized
by bishop Fridrek’
b. Og er þetta mál var rannsakað af lögmönnum
and when this case was investigated by lawyers
‘and when lawyers investigated this case’
c. Óttar gerði sem honum var boðið af Sighvati
Ottar did as him was ordered by Sighvat
‘Ottar did what Sighvat ordered him’
Using this method, we found at least 9 indisputable
examples of full DP Object Shift. This may not be the
exact number of such sentences in our corpus. First, in
addition to these examples, there are some borderline
cases, which may or may not be interpreted as instances of
Object Shift. Second, our searching method does not
guarantee 100% recall, as explained above. However, this
doesn’t really matter for our purposes. We have shown
conclusively that full DP Object Shift existed in Old
Norse, contrary to what has previously been claimed in
the literature; and we have demonstrated the efficiency of
our searching method.
Thus, our searching method has enabled us to strengthen
the evidence for the existence of derivational passive in
Old Norse.
5.
Conclusion
In this paper, we have demonstrated that it is possible to
use a tagging model trained on Modern Icelandic texts to
facilitate tagging of Old Norse narrative texts. By using
this method, we are able to tag a large corpus of Old Norse
with acceptable accuracy in a relatively short time – only
a fragment of the time it would have taken to build a
tagging model for Old Norse from scratch.
Furthermore, we have shown that a corpus tagged using a
rich tagset based on morphosyntactic features can fruitfully be used in the search for a number of syntactic constructions, and hence is a valuable tool in studying
syntactic variation and change. Of course, a morphologically tagged corpus like the one we have built doesn’t
amount to a fully parsed corpus. Several syntactic features
cannot be searched for using our method. However, given
the tremendous effort it would take to build a parsed corpus of this size, we think our method is an alternative that
must be taken seriously.
Later this year, we intend to make the tagged Old Norse
texts available on the web using the Xaira program
(www.oucs.ox.ac.uk/rts/xaira/) from the British National
Corpus. This will enable users to search the corpus for
complex patterns using both words and tags in the search
text. Thus, the corpus will hopefully be of great use to
anyone studying Old Norse language, literature, and culture.
4.2 Passive
Another controversial feature of Old Norse syntax is the
nature of the passive. It has sometimes been claimed
(Dyvik, 1980; Faarlund, 1990) that all passive sentences
in Old Norse are lexical but not derived by NP-movement
(or chain-formation). This claim has been disputed, for
instance by Benediktsson (1980), and it has been claimed
that the existence of agentive prepositional phrases
(by-phrases) would be an argument against this analysis,
since such phrases presuppose a derivational analysis of
passive sentences (Rögnvaldsson, 1995).
Be that as it may, it is quite clear that agentive prepositional phrases in passives are rather rare in Modern Icelandic, and hence, one would not expect to find many of
them in Old Norse. Faarlund (2004), for instance, quotes
two such examples but concludes: “This is very rarely
found, however.”
It is not easy to search for such examples in an unannotated electronic text. One would have to search for the
preposition af ‘by’, but this preposition is one of the most
frequent words in Old Norse so this search would return
thousands of sentences. However, once we have a
morphologically tagged text, it is relatively easy to search
for agentive prepositional phrases. We can search for a
past participle, followed by af, followed by a nominal
(noun, pronoun, adjective) in the dative. Since the
distinction between past participle forms and adjectives in
the neuter singular is not always clear, and the tagger
makes a number of errors in this classification, we also
search for the adjectives in addition to the past participles.
This search returns some 130 sentences. Most of them are
not instances of agentive phrases, since the preposition af
can also have other functions. Nevertheless, we have
found at least 15 sentences with agentive prepositional
phrases, only a few of which have previously been quoted
6.
Acknowledgements
This project was partly supported by a grant from the
University of Iceland Research Fund to the project “The
syntactic use of Old Icelandic POS tagged texts”.
Thanks to the Árni Magnússon Institute of Icelandic
Studies for providing access to the IFD corpus used in this
research.
Thanks are also due to three anonymous reviewers who
made many valuable comments which helped us improve
the paper considerably, although we have not been able to
follow all their suggestions.
45
7.
References
Kings of Norway]. Mál og Menning, Reykjavik, Iceland.
Kroch, A., Santorini, B., Delfs, L. (2004). Penn-Helsinki
Parsed Corpus of Early Modern English. http://www.
ling.upenn.edu/hist-corpora/PPCEME- RELEASE-1/
Kroch, A., Taylor, A. (2000). Penn-Helsinki Parsed Corpus of Middle English, second edition. http://www.
ling.upenn.edu/hist-corpora/PPCME2- RELEASE-2/
Kroch, A., Taylor, A. (2001). Verb-Object Order in Early
Middle English. In Pintzuk, S., Tsoulas, G., Warner, A.
(Eds.), Diachronic Syntax: Models and Mechanisms.
Oxford University Press, Oxford, UK, pp. 132-163.
Kroch, A., Taylor, A., Ringe, D. (2000). The Middle
English Verb-Second Constraint: a Case Study in Language Contact and Language Cange. In Herring, S.,
Schoesler, L., van Reenen, P. (Eds.), Textual Parameters in Older Language. John Benjamins, Philadelphia,
pp. 353-391.
Marcus, M.P., Santorini, B., Marcinkiewicz, M.A. (1993).
Building a Large Annotated Corpus of English: The
Penn Treebank. Computational Linguistics, 19(2), pp.
313-330.
Mason, L. (1999). Object Shift in Old Norse. MA thesis,
University of York, York, UK.
Megyesi, B. (2002). Data-Driven Syntactic Analysis –
Methods and Applications for Swedish. Doctoral
dissertation, Department of Speech, Music and Hearing,
KTH, Stockholm, Sweden.
Pind, J. (Ed.), Magnússon, F., Briem, S. (1991). Íslensk
orðtíðnibók [Icelandic Frequency Dictionary, IFD]
Orðabók Háskólans, Reykjavík, Iceland.
Rögnvaldsson, E. (1990). Orðstöðulykill Íslendinga sagna
[The Concordance to the Icelandic Family Sagas].
Skáldskaparmál, 1, pp. 54-61.
Rögnvaldsson, E. (1995). Old Icelandic: A Non-Configurational Language? NOWELE, 26, pp. 3-29.
Rögnvaldsson, E. (1996). Word Order Variation in the VP
in Old Icelandic. Working Papers in Scandinavian
Syntax, 58, pp. 55-86.
Rögnvaldsson, E. (2005). Setningafræðilegar breytingar í
íslensku. [Syntactic Changes in Icelandic.] In Thráinsson, H. (Ed.) Setningar. Handbók um setningafræði
[Sentences: A Handbook on Syntax]. (Íslensk tunga
III.) Almenna bókafélagið, Reykjavík, Iceland, pp.
602-635.
Rögnvaldsson, E. (2006). The Corpus of Spoken Icelandic and Its Morphosyntactic Annotation. In
Henrichsen, P.J., Skadhauge, P.R. (Eds.): Treebanking
for Discourse and Speech, Proceedings of the
NODALIDA 2005 Special Session on Treebanks for
Spoken Language and Discourse. Copenhagen Studies
in Language 32. Samfundslitteratur, Copenhagen,
Denmark, pp. 133-145.
Sundquist, J.D. (2002). Object Shift and Holmberg’s
Generalization. In Lightfoot, D. (Ed.), Syntactic Effects
of Morphological Change. Oxford University Press,
Oxford, UK, pp. 326-347.
Thráinsson, H. (2007). The Syntax of Icelandic. Cambridge University Press, Cambridge, UK.
Benediktsson, H. (1980). The Old Norse Passive: Some
Observations. In Hovdhaugen, E. (Ed.), The Nordic
Languages and Modern Linguistics 4. Universitetsforlaget, Oslo, Norway, pp. 108-119.
Bjarnadóttir, K. (2002). The Icelandic µ-TBL Experiment:
Preparing the Corpus. Paper presented at NLP1 final
session, January 9. GSLT, Växjö, Sweden.
Brants, T. (2000). TnT – A Statistical Part-of-Speech
Tagger. In Proceedings of the 6th Applied NLP Conference, ANLP-2000, Seattle, WA, pp. 224-231.
Degnbol, H. (1985). Hvad en ordbog behøver – og andre
ønsker [What a Dictionary Needs – and Others Wish
for]. In The Sixth International Saga Conference.
Workshop Papers I. Det arnamagnæanske institut,
University of Copenhagen, Copenhagen, Denmark, pp.
235-254.
Dyvik, H. (1980). Har gammelnorsk passiv? [Does Old
Norse have the Passive?] In Hovdhaugen, E. (Ed.), The
Nordic Languages and Modern Linguistics 4.
Universitetsforlaget, Oslo, Norway, pp. 82-107.
Faarlund, J.T. (1990). Syntactic Change. Toward a Theory
of Historical Syntax. Mouton, Berlin, Germany.
Faarlund, J.T. (2004). The Syntax of Old Norse. Oxford
University Press, Oxford, UK.
Halldórsson, B., Torfason, J., Tómasson, S., Thorsson, Ö.
(Eds.). (1985-86). Íslendinga sögur [The Icelandic
Family Sagas]. Svart á Hvítu, Reykjavik, Iceland.
Haugan, J. (2001). Old Norse Word Order and Information Structure. Doctoral dissertation, NTNU, Trondheim, Norway.
Helgadóttir, S. (2005). Testing Data-Driven Learning
Algorithms for PoS Tagging of Icelandic. In Holmboe,
H. (Ed.), Nordisk Sprogteknologi. Årbog 2004. Museum Tusculanums Forlag, University of Copenhagen,
Denmark, pp. 257-265.
Helgadóttir, S. (2007). Mörkun íslensks texta [Tagging
Icelandic Text]. Orð og tunga, 9, pp. 75-107.
Holmberg, A. (1986). Word Order and Syntactic Features
in the Scandinavian Languages and English. Doctoral
dissertation, University of Stockholm, Stockholm,
Sweden.
Holmberg, A., Platzack, C. (1995). The Role of Inflection
in the Syntax of Scandinavian Languages. Oxford
University Press, Oxford, UK.
Hróarsdóttir, Þ. (2000). Word Order Change in Icelandic:
from OV to VO. John Benjamins, Amsterdam, The
Netherlands.
Hwa, R., Resnik, P., Weinberg, A., Cabezas, C., Kolak, O.
(2005). Bootstrapping Parsers via Syntactic Projection
across Parallel Texts. Natural Language Engineering,
11(3), pp. 311-325.
Kristjánsdóttir, B., Halldórsson, B., Sigurðsson, G.,
Grímsdóttir, G.Á., Ingólfsdóttir, G., Torfason, J.,
Tómasson, S., Thorsson, Ö. (Eds.). (1988). Sturlunga
saga [The Sturlunga Collection]. Svart á Hvítu,
Reykjavik, Iceland.
Kristjánsdóttir, B., Halldórsson, B., Torfason, J., Thorsson , Ö. (Eds.). (1991). Heimskringla [The Sagas of the
46
Building a large dictionary of abbreviations for named entity recognition in
Portuguese historical corpora
1
Oto Vale, 2Arnaldo Candido Jr.*, 2Marcelo Muniz, 1Clarissa Bengtson, 1Lívia Cucatto, 1Gladis
Almeida, 4Abner Batista, 4Maria C. Parreira, 3Maria Tereza Biderman, 2Sandra Aluísio
Federal University of São Carlos (UFSCar), University of São Paulo (USP), State University of São Paulo (UNESP)
UFSCar - Department of Linguistics - Via Washington Luís, Km. 235 – C.P. 676 - CEP 13.565-905 - São Carlos, SP Brazil
2
USP - Centre of Computational Linguistics (NILC)/ Department of Computer Sciences
C.P. 668 - CEP: 13560-970 - São Carlos, SP, Brazil
3
UNESP, FCLAR - Department of Linguistics - Rod. Araraquara - Jaú Km1, Bairro dos Machados, C.P. 174, CEP: 14.800-901,
Araraquara, SP, Brazil
4
UNESP, IBILCE - Department of Linguistics - Rua Cristóvão Colombo, 2265, Bairro: Jardim Nazareth, CEP: 15054-000
São José do Rio Preto, SP, Brazil
E-mail: [email protected], [email protected], [email protected], [email protected],
[email protected], [email protected], [email protected], [email protected],
[email protected], [email protected]
1
Abstract
Abbreviated forms offer a special challenge in a historical corpus, since they show graphic variations, besides being frequent and ambiguous.
The purpose of this paper is to present the process of building a large dictionary of historical Portuguese abbreviations, whose entries include
the abbreviation and its expansion, as well as morphosyntactic and semantic information (a predefined set of named entities – NEs). This
process has been carried out in a hybrid fashion that uses linguistic resources (such as a printed dictionary and lists of abbreviations) and
abbreviations extracted from the Historical Dictionary of Brazilian Portuguese (HDPB) corpus via finite-state automata and regular
expressions. Besides being useful to disambiguate the abbreviations found in the HDBP corpus, this dictionary can be used in other projects
and tasks, mainly NE recognition.
1.
already being coined on this side of the Atlantic. On the one
hand, the speakers of those days faced a world materially
and culturally different from what was known in Europe;
therefore, they needed to designate referents of this new
universe, which were hitherto unnamed, using words from
the Portuguese linguistic system. The hundreds of native
languages then spoken in Brazil had their own vocabulary
for designating elements of the Brazilian fauna and flora, but
these words did not exist in European Portuguese. On the
other hand, customs and institutions gradually began to form
in this new society with the infusion of new cultures,
resulting in new words, different from those used in the
Portuguese metropolis.
To build the corpus, we collected documents in public
archives and libraries all over Brazil and in Portugal. This
corpus totals 2,458 texts; 287,570 sentences; 16,505,808
tokens (of which 368,850 are different from each other);
7,492,473 simple forms 1 (of which 368,529 are different
from each other); and 82.2 MB. In a similar endeavor related
to European Portuguese, researchers of the Universidade
Nova de Lisboa have built the "Corpus Informatizado do
Português Medieval" 2 , comprising Latin-Romance texts
from the ninth to the twelfth centuries, and Portuguese texts
from the twelfth to the sixteenth centuries, totaling some 2
million words. Our corpus was built to be processed with
corpus processing system UNITEX3 (Unicode – UTF-16)
Introduction
The Historical Dictionary of Brazilian Portuguese (HDBP),
the first of its kind, is based on a corpus of Brazilian
Portuguese texts from the sixteenth through the eighteenth
centuries (including some texts from the beginning of the
nineteenth century). The HDBP is a three-year project,
which started in 2006, developed under the sponsorship of
CNPq, Brazil. Organizing this historical dictionary has
required an extensive, time-consuming analysis of
documents, published texts and manuscripts produced by
eyewitnesses in the early stages of Brazilian history. One
important difficulty in compiling this corpus derived from
the absence of press agencies in colonial Brazil, which had a
precarious communication system. Only in 1808, after
escaping from Napoleon’s army, did the Portuguese
monarchy transfer the government of the Portuguese empire
to Brazil and improved communications. Moreover,
peculiarities affecting language must be considered, such as
biodiversity and multifaceted cultural traditions from
different regions of the country. To implement the HDBP
project, we created an integrated network of researchers
from various regions of Brazil and Portugal, including
linguists and computer scientists from 11 universities. Our
team comprises 18 researchers holding a PhD, with
complementary skills and expertise, and 23 graduate and
undergraduate students.
This project fills a gap in Brazilian culture, for it is
developing a dictionary that describes the vocabulary of
Brazilian Portuguese in the beginning of the country’s
history. At that time, Brazilian language was still dependent
on European Portuguese, even though some vocabulary was
*
Scholarship CNPq, Brazil.
This is the total number of words in the corpus that are
composed of letters belonging to Historical Portuguese alphabet.
2
Digital
Corpus
of
Medieval
Portuguese:
http://cipm.fcsh.unl.pt.
3
http://www-igm.univ-mlv.fr/~unitex/
1
47
and with Philologic4 (Unicode – UTF-8), since the latter is
web-based and includes several corpus-processing tools, as
for example AGREP 5 , used to check for similar or
alternative spellings in Philologic. To process this large
corpus, we have faced the typical problems researchers are
likely to encounter when dealing with old documents,
starting with text digitalization. Rydberg-Cox (2003) and
Sanderson (2006) state that, in historical Latin, Greek and
English texts, to mention just a few languages, words broken
at the end of a line are not always hyphenated; word-breaks
are not always used; common words and word-endings are
abbreviated with non-standard typographical symbols;
uncommon typographical symbols pervade non-abbreviated
words; and spelling variation is common even within the
same text. We encountered these same problems in the
HDBP project. First of all, the non-existence of an
orthographical system in the afore-mentioned centuries
generated a Babel of graphic systems being used by many
different scribes or copyists. Giusti et al. (2007) focus on this
difficulty introducing both an approach based on
transformation rules to cluster distinct spelling variations
around a common form, which does not always correspond
to the orthographic (or modern) form, and choices made to
build a dictionary of spelling variants of Brazilian
Portuguese based on such clusters.Another problem was
scribes’ habit of abbreviating words to facilitate handwriting
– there are many thousands of such abbreviations. Therefore,
for the correct understanding of texts, it was necessary to
expand abbreviations, a task that presents two main
difficulties. The first is related to the use of modern
knowledge sources to perform expansion, since gazetteers,
encyclopedias and heuristics currently in use do not address
directly the needs of historical material describing people,
places, and other entities that often do not appear in modern
sources (Crane & Jones, 2006). The second, and perhaps
most important, is that even if we had adequate knowledge
sources for expanding abbreviations they are highly
ambiguous with respect to meaning, which is critical for
understanding correctly not only the abbreviations
themselves but also the whole text (Kerner et al., 2004).
In general, if abbreviations are not expanded correctly they
can limit the effectiveness of: i) information extraction and
retrieval systems in digital libraries; ii) electronic index
creation from a corpus; iii) Natural Language Processing
(NLP) tools, such as taggers, parsers and named entity
recognition (NER) systems to enrich corpora linguistically.
Within the scope of the HDBP project, the failure of proper
abbreviation expansion hinders the correct editing of
dictionary entries. However, expanding each and every
abbreviation manually in a several million-word corpus is a
time-consuming, expensive and difficult – if not impossible
– task, due to the inherent ambiguity of noun abbreviations,
for example. In Section 4, we discuss our approaches to this
problem. Automatic acronyms and abbreviations
disambiguation have been given close attention in medical
and biomedical domains, since text normalization is an
important task for successful information retrieval and
4
5
extraction from texts in these areas (Pakhomov, 2002; Hong
et al., 2002; Schwartz & Hearst, 2003; Dannélls, 2006).
However, most of this automatic abbreviation
disambiguation research has focused on modern scientific
material, whereas historical corpora and digital libraries
remain largely ignored (Rydberg-Cox, 2003). Moreover,
NER systems have only begun to be implemented for digital
libraries (Crane & Jones, 2006). Taking the above into
consideration, the purpose of this paper is to present our
ongoing work to build a large dictionary of abbreviations
that contains the pair abbreviation and its expansion,
together with morphosyntactic and semantic information.
We are developing this process in a hybrid fashion, using
linguistic resources (such as a digitalized printed dictionary
of abbreviations and the authoritative lists of abbreviations
that accompany the material that is being digitalized) and
abbreviations extracted from the HDBP corpus via
finite-state automata and regular expressions. Since
expanding abbreviations is a costly process, these automata
were created to recognize larger patterns of abbreviations
that are NEs or the same pattern that has different types of
NEs. For example, the dictionary semantic tags allowed us
to identify new NEs, i.e., sequences of words that can be
identified as personal names in some contexts, but that are
categorized as place, river or organization names in other
contexts. We have been working on an iterative process,
which started with linguistic and common sense knowledge,
to attribute initial NE categories to an abbreviation and later
try to capture new NEs to update the respective dictionary
entry. Both the new NE categories and the larger
abbreviations gathered in this iterative process will be
inserted in our dictionary of abbreviations, which will be
useful for other projects and tasks, mainly named entity
recognition and abbreviation disambiguation in the HDBP
corpus. In the next section, we explain the details of the
HDBP corpus and the graphic form of abbreviations found
in it, as well as some historical corpus projects that addressed
the same issue. In Section 3, we describe the process of
building a dictionary of abbreviations. In Section 4, we
consider possible applications of the dictionary in the HDBP
project itself and in other scenarios: abbreviation lookup and
expansion; search for spelling variations of abbreviations;
and linguistic research. Section 5 contains our conclusions
and final remarks.
2. The HDBP corpus and its abbreviations
The texts in the HDBP corpus were written by Brazilian
authors and Portuguese authors who have lived in Brazil for
a long time. Among the texts selected for our corpus, there
are, for instance, letters of Jesuit missionaries, documents of
the bandeirantes (members of the exploratory expeditions
that pushed Brazilian borders far into inland areas), reports
of sertanistas (explorers of Northeastern Brazil), and
documents of the Inquisition. Table 1 shows more details
about the composition of our corpus.
Since the emphasis is on word meaning, we have selected
mainly published texts with minor editing. Examples of such
editing are the separation of words that come together in the
http://philologic.uchicago.edu/index.php
http://www.tgries.de/agrep/
48
original text, the introduction of punctuation marks,
paragraph mark-up to facilitate reading, and the insertion of
letters and words in places where editors were sure (or
almost) that such items were missing.
Data
XVIth XVIIth XVIIIth XIXth
Texts (%)
6.24
26.39
59.78
7.59
Sentences (%)
6.30
18.32
64.34
11.04
Simple Forms (%) 7.60
20.18
62.57
9.65
Megabytes (%)
7.23
19.95
63.09
9.73
(Friburger, 2002). In the historical corpus, this problem is
magnified by a large variation in the use of punctuation and
capitalization, which also affects the recognition of named
entities, since the corpus contains capitalized common words,
as if they were proper names, and proper names in lower
case. Table 2 illustrates problems related to abbreviations:
ambiguity and variants. The first column shows 13 different
expansions for the abbreviation “A”. The second column
illustrates 13 different forms of abbreviating the name of the
famous Brazilian city “Rio de Janeiro” (some of them in
lower case), which makes them hard to memorize.
Rio de Jan.ro
alteza (highness)
Rio de Janro
alvará (warrant)
Rio de Janr.o
Amaro (proper name)
Rio de Jan.o
Ana (proper name)
Rio de Jnro
anima (cheers up)
Rio de janro
ano (year)
Rio de janro
anos (years)
R o de jano
Antônio (proper name)
arroba (measure of weight, singular) R o de Janro
arrobas (measure of weight, plural) R o de janero
R o de Janro
Assembléia (assembly)
R o de Jnro
assinado (signed)
Rio de Janro
Atual (current)
Table 1: Distribution of texts by century
This decision was made to avoid potential problems during
corpus compilation; however, we still had to deal with the
following issues: 1) guaranteeing consistent assignment of
Unicode characters in the texts, since digitalization and OCR
correction have been done by different groups
geographically distant from one another; 2) treating the
graphic variation that alters frequency counts in the corpus,
thus causing difficulties for the selection of variants in
dictionary entries; and 3) expanding the abbreviations that
pervade the texts.
Table 2: Ambiguity and spelling variation in abbreviations
Most previous work on Brazilian Portuguese historical
corpus expands abbreviations manually, such as the project
“Para uma História do Português do Brasil”6 and “Projeto
Programa para a História da Língua Portuguesa”
(PROHPOR7). Also, the Tycho Brahe Project8 (Paixão de
Sousa & Trippel, 2006), whose purpose is to model the
relation between prosody and syntax in the process that led
from Classical to Modern European Portuguese, contains
tagged and parsed texts written by Portuguese authors born
between the sixteenth and nineteenth centuries. These texts
had their abbreviations expanded manually to facilitate
tagging and parsing. Although this corpus is large for the
task of syntactic analysis – it is currently composed of 46
texts and still growing – it remains manageable by manual
markup made with widely available standards in XML. The
large-scale Germany-wide project Deutsch.Diachron.Digital
(DDD) (Dipper et al., 2004) was set to build a diachronic
corpus of German with texts from the ninth century (Old
High German) to the present (Modern German) for linguistic,
philological and historical research. This is a long-term
project – it is planned to run over seven years – and its large
core corpus will reach 40 million words. The abbreviations
found in it will be expanded and annotated, based on
well-accepted international standards in XML.
All projects mentioned above expand their abbreviations
manually; however their development contexts differ from
that of HDBP, which has only three years to develop both a
large corpus and a dictionary. This is the reason why we had
Figure 1: Excerpt from Cartas Remetidas para Lisboa em 12
de julho de 1715 In: Negócios coloniais, L. Lisanti (ed.)
There are several graphic forms for the abbreviations found
in the HDBP corpus, some of which are shown in Figure 1:
a) abbreviations with a dot followed by a superscript piece of
text, as in “Janr.o”/Janeiro (January) and “corre.te”/corrente
(current), which appear in the first line;
b) abbreviations followed by a dot, as in “porq.”/porque
(because) and in the three instances of “q.”/que
(next/who/next, respectively) in the second and third lines.
To be consistent, we used the character “^” to denote
superscript, thus generating the forms “Janr.^o” and
“corre.^te” showed in (a) above, which can be processed
computationally. The same symbol is used when the
abbreviation does not possess a dot, but has a superscript
chunk, as in “O sor Jesus xpo”/O Senhor Jesus Cristo (The
Lord Jesus Christ), leading to the forms “s^or” and “xp^o”.
Other abbreviations display numerals, e.g., “8.bro”/Outubro
(October), or other characters, e.g., “@” for the word ano
(year). Some abbreviations only omit letters, as in
“Glo”/Gonçalo (proper name Gonçalo), “Jão”/João (proper
name João), “ldo”/licenciado (licensed), “Ros”/Rodrigues
(proper name Rodrigues), and “snr” or “snro”/senhor (sir).
An additional difficulty posed by abbreviations is that they
hinder the correct segmentation of sentences in UNITEX
6
“For a History of Brazilian Portuguese”:
http://www.letras.ufrj.br/phpb-rj/
7
“Project Program for a History of Portuguese Language”:
http://www.prohpor.ufba.br/projetos.html
8
http://www.ime.usp.br/~tycho/
49
to approach the problem related to abbreviation expansion in
a different way, detailed in Section 3.
employed lexicons together with corpus processing tools,
especially to expand and enrich a digitalized printed
dictionary (Flexor, 1991) with information about the NE
categories appearing in the HDBP corpus. Flexor (1991) is a
large alphabetically organized dictionary of abbreviations
from the sixteenth through the nineteenth centuries.
Although it has a large number of abbreviations (see Tables 3
and 4), most of them are not found in our corpus (only 16%
appear in the HDBP corpus). We performed an experiment to
recover abbreviations from the HDBP corpus using three
simple heuristics, to estimate the amount of abbreviations in
the corpus that is not present in the Flexor dictionary. We
found out 7,045 abbreviations with the heuristics; only 35%
of them (2,473) are in the Flexor dictionary. However, the
Flexor dictionary is still worth using as it has abbreviations
expansion. This dictionary is being revised to eliminate
entries that could be considered spelling variants, as in the
following example (pairs are composed of abbreviation and
expansion): (Bẽs, bens); (Bẽz, bens); (Bãda, banda), since the
tilde was part of the writing system of historical Portuguese.
Simple and Multi-word Abbreviations by Century
3. Building a dictionary of abbreviations
In recent years, NLP researchers have focused on
standardizing methods to construct linguistic resources,
which led to the development of tools now accepted
internationally. One of these construction standards, DELA
(Dictionnaires éléctroniques du LADL), was developed at
LADL (Laboratoire d’informatique documentaire et
linguistique, University of Paris 7, France), jointly with the
corpus-processing tool INTEX (Silberztein, 2000). DELA
became the standard tool for developing electronic lexicons in
the research network Relex9. These lexicons are used with
INTEX, and now also with its open-source counterpart
UNITEX (Paumier, 2006). This format allows for declaring
simple and compound lexical entries, which can be associated
with grammatical information and inflection rules. These
dictionaries are linguistic resources specifically designed to
perform automatic text processing operations. Types of
DELA are DELAF, which comprises inflected simple words,
DELAC and DELACF, for non-inflected and inflected
compound words, respectively. The dictionaries of simple
words (DELAS and DELAF) are lists of simple words
associated with grammatical and inflectional information.
The grammatical information is mainly morphological and
corresponds to gender, number, degree, case, mood, tense,
and person. However, with this format, it is possible to add
syntactic and semantic information gradually (Ranchhod,
2001). DELAF lexical entries have the following general
structure:
(Inflected
word),(canonical
form).(part
of
speech)[+(subcategory)]:morphological features
Types
Flexor
Flexor (%)
Intersection of
Flexor and
Corpus
Intersection of
Flexor and
Corpus (%)
Coverage (%)
3.1 Customizing UNITEX
XVIth XVIIth XVIIIth XIXth Total
2,050
4,091
14,376
9,939 21,869
9.37
18.70
65.74
45.45 139.26
754
1,323
2,447
1,710
21.37
37.49
69.34
48.46 176.65
3,529
16.13
Table 3: Abbreviations from Flexor (1991) by century,
showing the % of forms found in the HDBP corpus10.
Processing lexicographical tasks in a corpus is easier when
computational lexicons are available, and that was the reason
why we adopted UNITEX in the HDBP project. UNITEX
supports several languages, including Portuguese.
Language-specific resources are grouped in packets referred
to as idioms. When the UNITEX-PB (Muniz et al., 2005) was
created, a lexicon for contemporary Brazilian Portuguese was
incorporated into it. However, due to the peculiarities of
historical texts, several changes had to be implemented and a
new idiom was created, named “Português Histórico”
(Historical Portuguese). These changes included characters
that are no longer used in Portuguese, such as the long s (ʃ)
and the tilde (~) over consonants. Some diacritical marks
differ from the ordinary diacritics currently used in
Portuguese, because the former can be placed over
consonants. For instance, an accent mark over “m̃” was
common in Historic Portuguese. The introduction of such
characters was made possible using Unicode when the text
was being compiled.
This hybrid approach to build a dictionary has already been
successfully used to develop a dictionary of anthroponyms
(Baptista, Batista and Mamede, 2006) and was adopted in the
HDBP project as well. Besides, we have employed the
authoritative lists of abbreviations found in the books we
digitalized.
Simple and Multi-word Abbreviations by n-grams
Types
1
2
3
4
5
6
Total
or +
Flexor
17,872 1,624 833 527 302 711 21,869
Flexor (%) 81.73 7.42 3.81 2.41 1.38 3.25 100.00
2
3,529
Intersection 3,237 234 33 18 5
of Flexor
and Corpus
Intersection 91.75 6.60 0.94 0.51 0.14 0.06 100.00
of Flexor
and Corpus
(%)
3.2 The hybrid process to build a dictionary of
abbreviations
Table 4: Abbreviations from Flexor (1991), by size
3.2.1 Printed resources
In order to build our dictionary of abbreviations, we
Thus far, we have digitalized and processed abbreviations
9
10
http://ladl.univ-mlv.fr/Relex/introduction.html
50
Note that abbreviations can happen in more than one century.
abbreviations under letter A, 814 are named entities (<ENT>)
and 548 have the tag <INIT>. 1789 were simple abbreviations.
There are also 430 multi-word abbreviations in letter A. From
the 488 simple abbreviations under letter B, 260 are named
entities (<ENT>) and 138 have the tag <INIT>. Only 107
were common abbreviations. Some entries classified as
<ENT> are <INIT> as well, such as “Barb^ro” (barber), a
family name and a pattern used to introduce this profession.
There are also 45 multi-word abbreviations in letter B, such as
“Bn^s Ay^s” (Buenos Aires) and “Brigad^ro Insp^or”
(Brigadeiro Inspetor/Inspector Brigadier). As for letter C,
from the 2187 simple abbreviations, 364 are named entities
and 853 have the tag <INIT>. There are also 510 multi-word
abbreviations in letter C. All the multi-word abbreviations
will be annotated later.
from Flexor (1991) and some of the authoritative lists of
abbreviations to be used in the UNITEX system. Initially, the
information we had to include in the entries, gathered from
printed resources, was just the abbreviation, its expanded
form, and the century in which the text had been written.
However, considering information retrieval, we soon found
out that the canonical form was also extremely important and
should be in the dictionary, as it is required in the DELA
format. Therefore, we added this information, and now a
search for the canonical form capitão (captain), for instance,
produces the following forms (nonexhaustive list): Capitão,
capitam, Capitaõ, cappitão, Cappitam, capitães, Capitãis,
capitans and the abbreviated forms (nonexhaustive list):
Capaens
Capams
Capans
Capens
Captm
es
ms
ns
s
Cap
Cap
Cap
Cap
Capão
Captens
Captes
Capaens
Capittes
Capp.
Capms
Capnes
Capns
Capns
Cappão
ães
es
tes
es
Capp
Capp
Capp
Capt
Cappnn
3.2.2 Generic patterns to extract different categories of
NEs for an abbreviation
The use of heuristics is efficient for extending lexicons of NEs,
such as in the search for words (or n-grams) that begin with a
capital letter that is not in the beginning of a sentence or in the
search for words followed by titles and forms of address. Thus,
heuristic rules allow for the identification of named entities.
However, the identification of some abbreviated NEs, such as
“V. M.” (Vossa Mercê/archaic Portuguese for “you”), is
difficult, because the dot that follows V makes the NE look
like the beginning of a sentence (“. M.”), and therefore
impossible to be retrieved using the heuristic rule mentioned
above12.
An experiment 13 was performed to investigate NEs in the
historical corpus, in order to extend and enrich the dictionary
of abbreviations. This experiment was carried out with the
dictionary of abbreviations described in Table 3. The three
lists of abbreviations for letters A, B, and C and some short
lists of abbreviations were first tagged with HAREM
categories. The NEs received the tag <ENT>, whereas all
entries received the tag <ABREV>. The tag <INIT> was
created to designate abbreviated collocations found at the left
of certain types of NEs, thus yielding three subcategories that
were not present in HAREM, viz. <TITULO> (for
jobs/professions and titles/positions), <PARENTE> (for
family relations), and <TRATAMENTO> (for forms of
address, since they are very pervasive in Flexor’s dictionary).
Besides, all ten HAREM categories were used to
subcategorize <INIT>.
First, using the dictionary of abbreviated forms, we performed
a search in the corpus for tag <ABREV> (rule 1), which
resulted in 1,795,519 occurrences. Several of them were not
abbreviated forms, but stopwords with similar formats. In
addition, several occurrences were actually orthographic
variants that looked like abbreviations, such as bom/bõ. This
prompted us to re-examine the list of abbreviations to remove
non-abbreviated variants of stopwords and abbreviated forms
of stopwords (we call this the pre-processing phase). For
instance, the prepositions “por” and “para” were abbreviated
as “p.”. However, this was also the abbreviation for padre
(priest). Re-examining the list led us to create the rule 2 for
searching forms in a UNITEX graph or using the following
regular expression to locate retrieve abbreviated forms or the
form
“p.”
preceded
by
determiners:
Our dictionary of abbreviations differs from its counterparts
developed in UNITEX, mainly in the use of a larger number
of attributes. The most important attributes that have been
added are: ABREV, used to denote abbreviation; SEC16,
SEC17, SEC18, and SEC19 to indicate the century to which
the lexical entry refers (information from Flexor (1991)) – the
century attribute appears only in some entries, since it was not
always possible to identify the period in which the
abbreviation was used; <ENT>, to denote a named entity (NE)
and the tag <INIT>, which is a collocation to extract certain
types of NE. Each NE receives additional attributes,
according to the category it belongs to. These categories were
established by a taxonomy proposed in the evaluation contest
of systems for recognizing named entities in Portuguese
(HAREM 11 ), organized by Linguateca. Among the ten
HAREM categories, we have employed nine of them except
OBRA (titles, man-made things). Figure 2 shows some lexical
entries in DELA format. In the first line of the Figure 2,
Brg^es is the form found in the corpus, Borges is the
canonical form (lemma), N (noun) is the part-of-speech tag
for the entry, ENT+PESSOA+ABREV+SEC19 are
additional attributes, and ms (masculine singular) is the
morphosyntactic tagging. We also included the expanded
form (Borges), which may differ from the canonical form in
some cases.
Brg^es,Borges.N+ENT+PESSOA+ABREV+SEC19:ms/Borges
Brag.,Braga.N+ENT+PESSOA+LOCAL+ABREV+SEC18:ms/Br
aga
Br^ça,Braça.N+ENT+VALOR+ABREV+SEC19:fs/Braça
7^bro,setembro.N+ENT+TEMPO+ABREV:ms/setembro
B^eis,bacharel.N+INIT+TITULO+ABREV:mp/bacharel
B.,beco.N+INIT+LOCAL+ABREV+SEC18:ms/beco
Bat^am,batalhão.N+INIT+ORGANIZAÇÃO+ABREV+SEC16:m
s/batalhão
Bas^tos,bastardo.N+INIT+PARENTE+ABREV+SEC19:ms/basta
rdos
Figure 2: Entry samples from the dictionary
We have already processed letters A, B, C and some of the
authoritative lists of abbreviations. From the 3051 simple
11
12
Note that sentence breaking was not performed in the corpus
preprocessing phase.
13
In this experiment, we have used UNITEX version 2.0 and set
UNITEX to find the longest matches in its searches.
http://poloxldb.linguateca.pt/harem.php?l=classificacao_v3_sem
51
<ABREV>+((o+ao+do+ho).p\.)
With this regular expression, all abbreviated forms in the
dictionary can be retrieved, plus the form “p.” preceded by
determiners, thus decreasing the number of abbreviated forms.
On applying the pre-processing cited above and the rule 2, the
number of abbreviations dropped to 804,939. Before applying
a search using tags, we tested the hypothesis that a significant
number of abbreviated forms were either an NE or were in the
vicinity of an NE. This test was carried out with the rule 3
depicted by:
(<ABREV>+((o+ao+do+ho).p\.)).(<MOT>+<MOT><MO
T>+<MOT><MOT><MOT>+<MOT><MOT><MOT><
MOT>)
applied to a search for abbreviated forms containing one to
four elements. The number of retrieved abbreviations was
469,640. Therefore, further strategies are necessary to identify
NEs, since we observed in our corpus that abbreviations tend
to be close to each other. We carried out another search using
tags, in which we replaced the tag <ABREV> by <INIT>
(rule 4):
(<INIT>+((o+ao+do+ho).p\.)).(<MOT>+<MOT><MOT
>+<MOT><MOT><MOT>+<MOT><MOT><MOT><
MOT>)
With the rule 4, the number of occurrences dropped to 22,196.
More than 50% were the abbreviated form “S.”, which stands
for “Saint”, and abbreviated forms of address such as “S. M.”
(Sua Majestade/His or Her Majesty), “S. A.” (Sua Alteza/His
or Her Highness). The names of saints, however, were
commonly other types of NE, not associated with PERSON.
In fact, they were abbreviations for names of places
(fazenda/farm,
arrayal/hamlet,
mosteiro/monastery,
aldeia/village, bairro/district, villa/village etc.), rivers
(Corgo/Brook, rio/river), organizations (mosteiro/monastery,
fortaleza/fortress). We can check this information looking for
such words at the left of the abbreviated form in the excerpt
shown in Figure 3. This analysis of abbreviations productivity
is useful for identifying and contextualizing new NEs. The
use of these new attributes allows for sophisticated searches in
the HDBP corpus and in other historical corpora. This is
important, because we intend to make this resource available
for research under request, since we cannot make it public due
to copyright issues. It will be possible, for instance, to search
for all NEs from the eighteenth century or for all NEs related
to persons in their abbreviated forms.
helps to get insight of the lexical patterning of historical
corpora. Moreover, it is more useful to retrieve ambiguous
classes of NEs, as illustrated in Figure 4 for the name of
saints, since all of them are under letter S. However, we
believe it is not efficient to enlarge the dictionary with new
and different entries of a given category and its
subespecifications. To perform this focused task, we are
applying the same process defined for REPENTINO14 , a
repository of NEs from modern Portuguese. This process is
run in six steps, but the last one was not applied, for we
adopted the NE taxonomy defined in HAREM: 1) choose a
category for which you intend to search examples of entities;
2) decide which is the most appropriate strategy to search for
the examples: a) by tag <INIT>, such as in Rio S. Francisco;
b) by context, such as in “localizado na XXX” (located at
XXX), which strongly suggests that “XXX” is a place; or c)
by discriminating suffixes (modern organizations have in
their names characteristic particles such as “Ltda.”/Ltd. or
“S.A”/Co.); 3) construct the respective pattern to be searched
in a given corpus processor or to act as an independent
program, and conduct the search; 4) validate manually the
obtained candidates, considering the intended category; 5)
include positive candidates in the repository; 6) if necessary,
create a new category/subcategory, thus expanding the
taxonomic classification system.
However, this process had to be adapted to historical corpora,
because they have a large number of abbreviations and
spelling variations related to both abbreviated words and
expanded words. The requirement to accept a new NE from
the corpus was that at least one of the components should be
in the abbreviated form. We could not adopt the requirement
of capitalization, since in historical corpora proper names are
not always capitalized. To illustrate the adaptations of this
procedure for retrieving new NEs from a corpus, we discuss
a case study about hydronyms – names of rivers, streams,
creeks, and brooks found in the HDBP. Flexor’s dictionary
(1991) contains 18 entries with the pattern Rio XXX/River
XXX, but eight of them refer to the city of Rio de Janeiro
(the rest are R^o da Ribr^a, R^o de Reg^o, R^o de S.
Fran^co, R^o dos Alm^das, R^o G^de, R^o G^re, R^o
Gdr^e, R^o Gr^de, R^o G^re e R^o P^do); there is nothing
about Creek XXX (or its variants, brooks, streams), so we
began with ten entries. The chosen search strategies were:
pattern formed by tag <INIT> and contexts
“naveg*”/navigate (on), that includes the several
conjugations of the verb to navigate. However, words tagged
as <INIT> could appear in their abbreviated or expanded
form and, besides, we would have to deal with spelling
variations and synonyms (see Table 5). To deal with spelling
variations, we adopted two resources: the HDBP dictionary
of spelling variants, created according to the SIACONF
methodology proposed in Giusti et al. (2007) 15 , which
employs 43 transformation rules to cluster variants under
one orthographic form, and the Philologic resource of
searching for similar patterns, which uses AGREP. The
HDBP dictionary of spelling variants has 18,082 clusters,
totaling 41,710 variants. In spite of producing false-positives,
Figure 3: Concordances retrieved from sentences that have
the abbreviated form “S.”
3.2.3. Specific patterns to extract new NEs of a given
category
The use of generic search patterns in corpus processing tools
shown in Section 3.2.2 is easily carried out by linguists and
14
15
52
http://poloclup.linguateca.pt/repentino/
Available at http://moodle.icmc.usp.br/dhpb/siaconf.tar.gz
AGREP helps to complete variants resulting from
SIACONF. To deal with synonyms of river, we used the
Brazilian Portuguese Electronic Thesaurus TEP (Greghi et
al., 2002).
Searching patterns
(63)
Sources
(5)
rio
river
Right
Occurrences
(112)
arroio, córrego, corrente, regato, synonyms
regueira, regueiro, ribeirão,
ribeiro, riacho, rio, veia, veio
arroyo, corrego, corego, corgo spelling
regueyro, ribeiraõ, ribeyrão,
variants
ribeyraõ, rybeirão, rybeyrão,
rebeirão, rebeyrão, ribeirao,
ribeiro, ribeyro, rybeiro, ribejro,
rjbeyro, rybeyro, riaxo, ryo, rjo,
rrio, veya, veyo
c^te, cor, cor^e, cor^te, corr^e,
corr^te, cort^e, crr^e, curr^te,
r^bro, r^o, r^ro, reb^o, rib.^ro,
rib^o, rib^ro, riber^o, ribr,
ribr^o, ryb^o, ryb^ro, rybr^o,
r^bro, r , r\.^o
abbreviations
from Flexor
(1991)
naveg*
context
categories. In the context of information retrieval for
historical documents, it may be useful to gather, for example,
texts reporting certain facts that happened in a certain place.
If the index is the expanded abbreviation, we can easily find
all abbreviations for a word, such as Bahia (a Brazilian city
in the northeastern coast of Brazil). Next, we can locate
passages in the corpus related to those abbreviations.
Our dictionary of abbreviations was designed to recognize
large patterns of complete abbreviations. It also includes a
specific tag for dealing with jobs/professions and titles and
forms of address, such as capitão (captain), frei (friar),
promotor
(prosecutor),
Ilustríssimo
(Most
Illustrious/Honorable), Dom (Don), Majestade (Majesty),
Senhor (Sir), and family relations, such as cunhada
(sister-in-law), primo (cousin). In linguistic research, it is
very important to know whom the text is talking about or
whom it is talking to. If we can determine the authorities that
are being addressed in a specific text, we can identify the
words used in that specific level of formality, given that a
letter written to an ordinary person does not contain the same
words and level of formality as one written to a monarch, and
this is possible because we used NEs and other specific tags.
79
13
7
11
2
Table 5: Searching patterns for hydronyms
The manual validation is the slowest step (we ckecked
27,808 ocorrencies in 160 minutes – 1,100 checkings per
hour), but easier in concordancers, since the pattern formed
by abbreviations stands out, which facilitates checking. As a
result of this case study we have now 122 abbreviations
under category LOCAL, specifically rivers and words
related to watercourses, displaying their morphology in this
semantic group in the HDBP corpus. Some examples are:
Ribeyrão de N. Sr.^a do Carmo/Ribeirão de Nossa Senhora
do Carmo; Corgo de S. Gonçalo/Córrego de São Gonçalo;
rib.^o do Tombadouro, ribeirão do Tombadouro; coRego
Ant.^o da Silua, Córrego Antonio da Silva; Rio M.^el Alves,
Rio Marechal Alves; R^o doce, Rio doce.
Figure 4: Search for pattern “b” in UNITEX; the program
Dicionário helps the manual disambiguation of the
abbreviation “b” (in the top right corner). All
subcategorization of NEs will be included later.
4. Applications of the dictionary of
abbreviations
5. Conclusions and Future Work
An example of use of the HDBP dictionary of abbreviations
is the application of UNITEX together with the software
Dicionário16 (Muniz et al., 2005) to assist lexicographers in
manually identifying possible canonical forms for an
abbreviation or for expanding an abbreviation. Using the
concordancer shown in Figure 4, a linguist may find
examples of abbreviations in the corpus, but may not be
aware of the possible expanded forms for a given
abbreviation. On using the software Dicionário together with
the concordancer, the abbreviations can be quickly identified
and associated with their possible canonical forms and
To sum up, the Historical Dictionary of Brazilian
Portuguese is not only a pioneer project, but also a
fundamental tool for recapturing and registering the
country’s early history through its vocabulary. The
compilation of a corpus of historical texts is therefore a
crucial step to achieve such aim, since it allows researchers
to retrieve the lexicon of a given period. The lexical,
morphological, syntactic, and typographic characteristics
identified in these texts have been the object of study of
various members of our team, which includes philologists,
linguists and computer scientists. Among the peculiarities of
historical texts, the abbreviated forms pose a special
challenge. In addition to their high frequency and ambiguity,
a researcher is also faced with the fact that, as far as historical
16
The software Dicionário is a Java application that handles any
dictionary compacted in the DELA format, and allows searching
for inflected words.
53
documents are concerned, there are no standard graphic
forms, and abbreviations reflect this inconsistency,
displaying a large number of variations. Taking this fact into
account and to make a lexicographer’s task feasible, special
attention was given to abbreviations. An electronic
dictionary of abbreviated forms is being built based on
printed resources, using the DELA format, which allows us
to categorize each new entry morphosyntactically,
semantically and pragmatically. New NE categories of
abbreviations were found using semantically categorized
abbreviations, UNITEX graphs and regular expressions to
examine the vicinity of abbreviated forms. Since the process
to expand abbreviations demands considerable expertise,
these automata and regular expressions were created to
recognize only larger patterns of abbreviations that are NEs,
spelling variations and synonyms of NEs or the same pattern
that has different types of NEs, given that context will
provide meaning. With regard to enlarging the dictionary of
abbreviations, we focused on a specific NE category (places),
subcategorizing it further (hydronyms). This experiment
provided us with evaluation data with regard to time spent
and productiveness rate of the semi-automatic approach we
decided to adopt to guarantee high accuracy for the
classification process. We concluded that this approach is
worth pursuing once we need to guarantee a high precision
classification. In the future, we intend to make this corpus
and the dictionary of abbreviations available for those
studies on history to which correct NE classification is
crucial and mainly as a resource for NE recognition systems.
Tours. Tours.
Giusti, R., Candido Jr, A., Muniz, M., Cucatto, L., Aluísio, S.
(2007). Automatic detection of spelling variation in
historical corpus: An application to build a Brazilian
Portuguese spelling variants dictionary. In Proceedings of
the Corpus Linguistics 2007 Conference, Matthew Davies,
Paul Rayson, Susan Hunston, Pernilla Danielsson (eds.).
Greghi, J. G., Martins, R. T., Nunes, M. G. V. (2002).
Diadorim: a Lexical database for Brazilian Portuguese In
Proceedings of the International Conference on Language
Resources and Evaluation LREC 2002, Manuel G.
Rodríguez and Carmem P. S. Araujo (Eds.), v. IV, pp.
1346--1350.
HaCohen, K. Y., Kass, A., Peretz, A. (2004). Baseline
Methods for Automatic Disambiguation of Abbreviations
in Jewish Law Documents. In EsTAL: International
Conference on Advances in Natural Language Processing
N. 4, Alicante. Lecture Notes in Computer Science. 3230.
Berlin: Springer, pp. 58--69.
Hong, Y., Hripcsak, G., Friedman, C. (2002). Mapping
abbreviations to full forms in biomedical articles. J Am
Med Inform Assoc. May–June, 9(3), pp. 262--272.
Muniz, M., Nunes, M. G. V., Laporte, E. (2005).
UNITEX-PB, a set of flexible language resources for
Brazilian. In III Workshop em Tecnologia da Informação e
da Linguagem Humana, pp. 2059--2068.
Paixão de Sousa, M. C., Trippel, T. (2006). Metadata and
XML standards at work: a corpus repository of historical
Portuguese texts. In Proceedings of V International
Conference on Language Resources and Evaluation
(LREC 2006), Genoa, Italy.
Pakhomov, S. (2002). Semi-supervised Maximum
Entropy-based Approach to Acronym and Abbreviation
Normalization in Medical Texts. In Medical Texts
Proceedings of ACL 2002. pp. 160--167.
Paumier, S. (2006). Manuel d’utilisation du logiciel
UNITEX. IGM, Université Marne-la-Vallée. Available at
6. Acknowledgments
The authors are grateful to CNPq (Brazil) for supporting this
research.
7. References
Baptista, J., Batista, F.; Mamede, N. (2006). Building a
Dictionary of Anthroponyms. In Computational
Processing of the Portuguese Language, 7th International
Workshop, PROPOR 2006, R. Vieira et al. (eds.)
Proceedings. Lecture Notes in Computer Science 3960,
Berlin: Springer, pp. 21--30.
Crane, G., Jones, A. (2006). The challenge of Virginia
banks: an evaluation of named entity analysis in a
19th-century newspaper collection. In Proceedings of the
6th ACM/IEEE-CS Joint Conference on Digital Libraries,
pp. 31--40.
Dannélls, D. (2006). Automatic Acronym Recognition. In
EACL 2006, 11th Conference of the European Chapter of
the Association for Computational Linguistics. pp.
167--170.
Dipper, S.; Faulstich, L.; Leser, U.; Ludeling, A. (2004).
Challenges in modelling a richly annotated diachronic
corpus of German. In Proceedings of the Workshop on
XML-based Richly Annotated Corpora. pp. 21--29.
Flexor, M. H. (1991). Abreviaturas - Manuscritos dos
Séculos XVI ao XIX. 2nd ed. São Paulo: UNESP. 468 p.
Friburger, N. (2002). Reconaissance automatique de noms
propres: Application à la classification automatique de
textes journalistiques. Thèse (doctorat). Université de
http://www-igm.univ-mlv.fr/~unitex/ManuelUnitex.pdf
Ranchhod, E. M. (2001). O uso de dicionários e de
autômatos finitos na representação lexical das línguas
naturais. In Tratamento das Línguas por Computador.
Uma Introdução à Lingüística Computacional e suas
Aplicações, E. Ranchhod (ed.), Lisbon: Caminho, pp.
13--47.
Rydberg-Cox, J. A. (2003). Automatic disambiguation of
Latin abbreviations in early modern texts for humanities
digital libraries. In Proceedings of JCDL, 03, pp.
372--373.
Sanderson, R. (2006). "Historical Text Mining", Historical
"Text Mining" and "Historical Text" Mining: Challenges
and Opportunities. Talk presented at the Historical Text
Mining Workshop, July 2006, Lancaster University, UK.
Schwartz, A. M., Hearst, M. (2003). A simple algorithm for
identifying abbreviation definitions in biomedical texts. In
Proceedings of the Pacific Symposium on Biocomputing
(PSB) 2003. pp. 451--462.
Silberztein, M. (2000). Intex: a FST toolbox. Theoretical
Computer Science, 231, pp. 33--46.
54
A Conversant Robotic Guide to Art Collections
D. Vogiatzis∗ , D. Galanis†, V. Karkaletsis∗, I. Androutsopoulos†‡ , C.D. Spyropoulos∗
∗
†
Institute of Informatics and Telecommunications, NCSR “Demokritos”
Athens, Greece
Department of Informatics, Athens University of Economics and Business
Athens, Greece
‡
Digital Curation Unit, Research Centre “Athena”
Athens, Greece
Abstract
We present the dialogue system of a robot that has been developed to serve as a museum guide. The robot interacts with human
visitors in natural language, receiving instructions and providing information about the exhibits. Moreover, being mobile, it physically
approaches the exhibits it provides information about. Although the robotic platform contains many modules, including navigation,
speech recognition and synthesis, our focus in this paper is the dialogue system, which supports the sessions between humans and the
robot, as well as the natural language generation engine, which generates the text to be spoken. Both modules are closely interwined and
depend on an ontology represented in OWL. The robot supports dialogues in both English and Greek.
1. Introduction
This paper presents the natural language interaction capabilities of a mobile robotic tour guide developed in
the Greek project Xenios.1 The robot adopts a simple
finite-state architecture for dialogue management, which is
complemented by multilingual natural language generation
from OWL ontologies; the latter is used to produce personalised descriptions of exhibits, currently in English and
in Greek. The robot guides the visitors to the premises of
the Hellenic Cosmos, the cultural centre of the Foundation
of the Hellenic World (FHW).2 The vision and navigation
components of the robot were developed by the Foundation
of Research and Technology Hellas.3
Robots are being deployed experimentally as tour guides in
science parks and museums (Bennewitz et al., 2005; Chiu,
2004; Thrun et al., 1999); however, the issue of natural language has been somewhat neglected. Most of the robotic
platforms acting as guides rely on prerecorded texts, which
are spoken by a speech synthesiser. In contrast, in our work
the focus is on dynamic text generation from domain ontologies, in the context of spoken dialogues. Spoken Dialogue Systems (SDS) typically involve speech recognition,
dialogue management, language interpretation, language
generation, and speech synthesis. In addition, they require
domain specific knowledge and linguistic resources (e.g.,
domain lexicons). CSLU4 , Dipper5 and Trindikit6 are three
representative examples of software platforms that support
the development of SDSs.
SDSs are divided into three major categories, depending on
who controls the dialogue (McTear, 2004). In system initia1
http://www.ics.forth.gr/xenios/description.html
http://www.fhw.gr/index en.html
3
http://www.forth.gr/
4
http://cslu.cse.ogi.edu/toolkit/
5
http://www.ltg.ed.ac.uk/dipper/
6
http://www.ling.gu.se/projekt/trindi/trindikit/
2
tive SDSs, the system asks a sequence of questions to elicit
the required parameters of the task from the user. In user
initiative SDSs, the dialogue is controlled by the user, who
asks the system questions in order to obtain information.
In mixed initiative SDSs, the dialogue control alternates
between the two participants; the user can ask questions
at any time, but the system can also take control to elicit
required information or to clarify ambiguous information.
In all SDSs, a dialogue management module is required to
determine what questions the system should ask, in what
order and when, as well as to anchor user utterances to
the discourse context (e.g., resolve referring expressions).
There are three main categories of dialogue management
techniques: finite state based, frame based, and plan based.
The most common category are the state based ones, which
use graphs. Dialogue is modeled as a network of states connected by edges. At each state, the system can perform one
of the following:
1. ask the user for specific information, possibly listing
expected answers (in effect offering menu options),
2. generate a response to the user, or
3. access an external application.
In Xenios, we use system initiative dialogues, with finite
state dialogue management, and particular options offered
to the user at each state. The advantages of this technique
are faster development and more robust systems, at the expense of limited flexibility in the dialogue structure.
Recent research has started to examine the efficient use
of domain ontologies in dialogue systems (Milward and
Beveridge, 2003). For instance, the domain specific lexicon and the grammar of the automatic speech recognition
(ASR) component can be partially derived from the ontology. Furthermore, the natural language generation component (NLG) can generate descriptions of the ontology’s instances or classes, as in our case. The ontology can also be
55
used during language interpretation, for example to locate
properties the user’s utterances may refer to.
An entire strand of work in Natural Language Generation
(NLG) has focused on producing textual descriptions of
classes and instances (entities) of domain specific ontologies (Reiter and Dale, 2000). A representative example is
ILEX (O’Donnell et al., 2001), which was mainly demonstrated with museum ontologies. In more recent work, a
multilingual extension of ILEX was developed in the context of the M - PIRO project (Isard et al., 2003; Androutsopoulos et al., 2007). The system was tested in several
domains including museum exhibits and computing equipment. Similar systems have also been built to automatically
produce medical reports (Bontcheva et al., 2004) and drug
descriptions (Bontcheva and Wilks, 2004).
The rest of this paper is organised as follows: the architecture of the dialogue system is presented in section 2; in
section 3, we present the resources that are used in the dialogue system; then, in section 4, we present the natural
language generation module; in section 5, we provide samples of a visitor’s interaction with the robot; the paper ends
with an overview and conclusions in section 6.
2. Architecture of the Dialogue System
The dialogue system consists of resources and modules (see
Fig. 1). The modules are the Dialogue System Manager
(DSM), the Natural Language Generation Engine (NLG),
the Automatic Speech Recognition (ASR) engine, the Text
To Speech (TTS) synthesiser and the Gesture Recogniser.
The latter can detect a set of three human gestures (yes, no,
quit), and it is not described here. As TTS we have used the
Demosthenes speech synthesiser (Xydas and Kouroupetroglou, 2001). As ASR we have used the BabEar system
from Acapela7 , which also provides natural language understanding facilities; the latter are based on a domain specific lexicon and a grammar.
The dialogue system manager (DSM) is the “actor” of the
whole dialogue system, in the sense that it is the module
that invokes and coordinates all the other modules. The
DSM initiates the dialogue, and expects the visitor to select an option from the ones offered. In order to decide
the next dialogue state and the text it will utter (through
the TTS unit), it takes into account the dialogue model, the
interaction history of the user, as well as the location of
the robot. All the above contribute into creating more natural dialogues. Finally, there is the communication server,
which enables the inter-module communication.
The resources are: the dialogue model; the resources of
the NLG, which are discussed separately below; user models, including databases recording the interaction history of
each user; and some databases that hold canned text and
other information used during the dialogue. The modules
are not domain specific; consequently, they can be easily
transferred to another domain, with the exception of the
ASR, which needs adjustments. On the other hand, the
resources are domain specific and must be designed from
scratch for each domain.
7
http://www.acapela-group.com/index.asp
3. Resources of the Dialogue System
Dialogue Model
The dialogue model is based on a finite state automaton.
At each state, the robot utters a particular question or statement. The transitions from a state represent the possible
options offered to the human at that state. The dialogue
model supports three sessions: visitor welcome, tour and
dialogue termination. The dialogue model was developed
especially for museums that host art collections, and are
equipped with show rooms, where informative movies can
be played.
In Fig. 2 we present an overview of the dialogue model.
The first session (topmost node), captures the part of the
dialogue session from the moment the visitor approaches
the robot, till the moment he validates his ticket (validation
occurs through a bar code reader, carried by the robot). The
bottommost node represents the termination of the dialogue
session. The second session (mapped to nodes 2, 3 and 4)
concerns the presentation of the buildings and programmes
of FHW to a visitor. The robot can talk about particular
programmes, offer tour guidance, or do both.
The middle level nodes do not correspond to individual
states; instead each one corresponds to approximately 10
states. Moreover, in each state there is a template which
combines canned text with calls to the NLG in order to produce the final text to be uttered by the TTS.
Domain Ontology
The ontology represents the buildings and rooms of FHW,
as well as the educational programs that are offered. The
information about the buildings and the rooms is essential to the robot’s navigation, to provide guidance instructions to visitors, and to relate educational programmes with
their physical location. For instance, the following ontology extract includes the class Area, as well as the entities
that belong to that class. For instance, first-dome and
ground-dome represent the first and second floor of the
dome (a semi-spherical movie projection hall).
Area
ground-dome
first-dome
screening-room
virtualReality-area
basement1-building1
basement2-building1
first-building1
ground-building1
The next ontology extract represents two types of programs
that are offered, in particular there are documentaries about
Miletus and the Black sea, as well as educational programs
about the Ancient Agora and the ancients’ contribution to
mathematics.
Programme
Documentary
miletus
black sea
Educational
ancientAgora
56
sources of the OWL ontology are exported by ELEON in
the RDF format the NLG module requires.
The ontology was authored by a group of people from
FHW. The group involved archaeologists that provided the
initial information in the form of a collection of documents; this collection had been used before to produce
paper leaflets and other promotional material. The same
group extracted the most important paragraphs from the
collection. The information of the paragraphs was then entered in ELEON, with additional information obtained from
an architectural plan. Furthermore, FHW edited the contents of the canned text database, by using a simple editor
developed for this purpose.
Figure 1: System Architecture
Figure 2: Dialogue Model
mathematics
The property takes-place-at, which is assigned to
every programme, takes as its values entities of the
Area class; for instance, <miletus, takes-place-at,
screening-room>. Summarising, the ontology describes
all the buildings and rooms where educational activities
take place, as well as a significant part of the programmes
that are currently offered.
The ontology can be updated by adding visiting areas, new
exhibits, or by updating information on already existing areas and exhibits. The use of the ontology enables the system to describe the newly added or updated objects without
further configuration. The ontology is authored in OWL,8
the Semantic Web standard language for specifying ontologies, using the ELEON editor (Bilidas et al., 2007). Furthermore, ELEON allows annotating the ontology with the
linguistic resources required by the NLG engine. User
types (e.g., adult, child, expert) can also be defined with
ELEON; we discuss user types below. The linguistic re8
http://www.w3.org/TR/owl-features/
User Types
User types such as adult, child, expert, are defined with
the aid of the ELEON authoring tool. The user types
are defined when creating the domain ontology, and do
not change afterwards. User types are quite useful, since
they permit extensive personalisation of the information
that users receive (Androutsopoulos et al., 2007). Thus,
user types determine the interest of the ontology facts (e.g.,
some facts about painting techniques may be too elementary for experts), as well as how many times a fact has to
be repeated before the system can assume that a user of a
certain type has assimilated it (e.g., how many times we
need to repeat the duration of the Roman period). In addition, user types specify the appropriateness of linguistic
elements (e.g., lexicon and microplans, see below), as well
as parameters that control the maximum desired length of
an exhibit description. Finally, different synthesiser voices
can be chosen for different user types. Orthogonal to the
user types is the multilingual support (English and Greek)
for the lexical items and the microplans.
Interaction History and Application Databases
The interaction history records information showing the exhibits each user has seen, the information the system has
conveyed, as well as the natural language expressions the
system has used to convey the information. This allows
the system to avoid repeating the same information and
expressions. The interaction history is kept in a database
that is controlled by the Personalisation Server, which can
also produce statistical data. Moreover, it can act as a recommender system (based on past usage data) upon the arrival of a new user (Paliouras et al., 2006). There is also a
Canned Text Database, which contains fixed text that will
be spoken at the commencement, at the end, or at an intermediate stage of the visitor’s interaction with the dialogue system. Canned texts also contain some string variables that are instantiated during the dialogue session. Finally, there is a Domain Specific Database, which in effect
contains instances of the ontology, for example, particular
buildings, programs and rooms. This information is extracted from the ontology that the NLG module uses (Galanis and Androutsopoulos, 2007).
4. The Natural Language Generation engine
The natural language generation engine of Xenios, called
NaturalOWL, is heavily based on ideas from ILEX and M 57
PIRO .9
Unlike its predecessors, NaturalOWL is simpler
(e.g., it is entirely template-based) and it provides native
support for OWL ontologies. Currently, the system supports English and Greek.
NaturalOWL adopts the typical pipeline architecture of
NLG systems (Reiter and Dale, 2000). It produces texts
in three sequential stages: document planning, microplanning, and surface realisation. In document planning,
the system first selects the logical facts (OWL triples)
which will be conveyed to the user and it specifies the document structure. In micro-planning, it constructs abstract
forms of sentences, then it aggregates them into more complex periods, and finally it selects appropriate referring expressions. In surface realization, the abstract forms of sentences are transformed into real text, and appropriate syntactic and semantic annotations can be added, for example
to help the TTS produce more natural prosody. The system is also able to compare the described entity to other
entities of the same collection (e.g., “Unlike all the vessels
that you saw, which were decorated with the black-figure
technique, this amphora was decorated with the red-figure
technique.”).
4.1. Document planning
Content Selection
When NaturalOWL is instructed to produce a description
of a class or an entity, it attempts to locate all the logical facts of the OWL ontology (OWL triples of the form
<S, Property, O>, where S and O are the semantic subject and object of the fact, respectively) that are relevant to
that entity or class. First, it selects all the facts that are directly relevant. For example, in the case of the entity whose
identifier is exhibit24, it first selects, among others, the
following facts, which associate exhibit24 with class
aryballos and the entities archaelogical-delos,
iraion-delos and archaic-period.
<exhibit24,rdf:type,aryballos>
<exhibit24,current-location,
archaelogical-delos>
<exhibit24,location-found,iraion-delos>
<exhibit24,creation-period,archaic-period>
Then, the system selects iteratively facts that are indirectly
relevant to the described entity or class. In our example, the
second iteration would add facts like the following:
<archaic-period,covers,
archaic-period-duration>
<aryballos,rdfs:subclassOf, vessel>
The final set of selected facts results by removing the already assimilated facts and then the facts with the lowest
interest. The interest of each fact and the number of times
a fact has to be mentioned in order to be considered as assimilated is specified in the user types (Androutsopoulos et
al., 2007).
9
See http://www.ltg.ed.ac.uk/methodius/ for information on
descendant of M - PIRO’s generator.
METHODIUS , another
Text Planning
The selected facts of the first iteration are ordered
based on a domain-specific partial order of their properties (e.g., current-location must be mentioned after
creation-period). More specifically, the partial order is
created with the help of the ELEON tool, which allows the
users to assign to each property an order score, an integer
from 1 to 10. Smaller order scores indicate that facts with
those properties should be mentioned earlier in the generated description. The order is stored in the RDF annotations of the ontology. The selected facts of the second or
subsequent iterations are placed immediately after the corresponding facts of the first iteration.
4.2. Micro-planning
Abstract sentence forms
The RDF annotations of the ontology map each OWL property to one or more natural language templates. The templates consist of a list of slots and instructions showing how
to fill them in. Each slot can be filled in with any of the following:
• A referring expression pointing to the semantic subject
of the fact.
• A fixed string. If the string is a verb, it is specially
tagged along with tense and voice. Prepositions are
also tagged. These tags are used in aggregation.
• A referring expression pointing to the semantic object
of the fact.
The final sentences are constructed by concatenating the
slot values. For example, for the current-location
property, the following microplan is defined:
Microplan(MicroName: "templ1",
ForProperty="current-location") ->
Slot(Type:String, Value: "today")
Slot(Type:Owner, Case: nominative)
Slot(Type:Verb, Value: "is exhibited",
Voice: Passive, Tense: present)
Slot(Type:Prep, Value: "in")
Slot(Type:Filler, Case: accusative)
If the above microplan is applied to the fact <exhibit24,
current-location, archeological-delos>, the
abstract sentence “today Ref(Entity: exhibit48,
Case: nominative) is exhibited in Ref(Entity:
archaelogical-delos, Case: accusative)” is
produced. The Ref tags represent the referring expressions
which will be produced in a following stage.
Aggregation
The produced abstract sentences are then aggregated into
longer sentences using a set of aggregation rules similar to M - PIRO’s (Melengoglou, 2002). For example, NaturalOWL’s rules aggregate phrases that share the same semantic subject (e.g., by adding the connective “and”), or
pairs of phrases where the semantic object of the first one
is the same as the semantic subject of the second one and
the second phrase’s verb is “to be”, as illustrated below.
58
This aryballos was found in the Heraion of
Delos. It is currently exhibited in the
Archaeological Museum of Delos.
->
This aryballos was found in the Heraion of
Delos and it is currently exhibited in the
Archaeological Museum of Delos.
This is an aryballos. An aryballos is a kind
of vessel.
->
This is an aryballos, a kind of vessel.
Referring expressions generation
In this stage, the system generates the appropriate referring
expressions. When an entity or class is introduced for the
first time, its name (e.g., “Bancroft Chardonnay is a . . . ”)
or the name of its class is used (e.g., “this is an aryballos”).
In all subsequent references, an appropriate pronoun is used
(e.g., “it was found in the Heraion of Delos”). If the focus
(semantic subject of facts) shifts to another entity or class,
again the name or the class of the new focus is first used,
followed by pronoun references. When the focus returns to
the original entity, a demonstrative (e.g., “this aryballos is
made of . . . ”) or the name of the original entity is used.
4.3. Surface Realization
Finally, in surface realization the system simply concatenates the slot values of the sentence specifications. In addition, syntactic or sematic markup can be added, as already
mentioned.
5. Sample Interaction
Below we present an excerpt from a dialogue between the
robot and a visitor. In the robot’s utterances, the words that
appear in italics are generated by the NLG module, whereas
the other words are canned texts, with some intervening values of variables. As it can be seen, each robotic utterance
is a mixture of dynamically generated and canned text, and
as the dialogue progresses, the robot traverses the states of
the dialogue model.
Robot Ok. Let us talk about the programmes of the Hellenic World. There are 6 programmes. The following:
a programme about the Ancient Agora of Athens, a
documentary about the Olympic Games, . . . . Please
tell me which programme you are interested in.
Human I am interested in the Ancient Agora of Athens.
R Right. This is an educational programme. Its title is:
“In the Ancient Agora of Athens”. It uses multimedia
technologies and it is hosted in building 1. Its subject is history and it was established in 2007. It is addressed to children from 9 to 15 years old. Would you
like me to guide you to the place of the programme?
The following excerpt illustrates a case where the robot
does not understand the user’s utterance:
R Which building would you like to visit? I remind you
that the Hellenic World has the following buildings:
building 1, building 5, building 7, Tholos, Research
Centre.
H The first one.
R I am sorry! I did not understand. I asked you which
building you would like to visit. The answers I expected were: building 1, building 5, building 7, Tholos and Research Centre. Please select one or use the
touch screen.
6.
Conclusions and Future Work
We have described the dialogue system of the Xenios robotic guide. The dialogue system comprises many modules and resources. The central module is the dialogue
manager that supports robot initiative dialogues. During
human-robot interaction, the robot’s utterances are a mixture of canned and dynamically generated text. Dynamic
text generation proceeds in three stages: document planning, microplanning, and surface realisation. One of the
most important resources of the system is the enriched domain ontology, which is authored by using the ELEON ontology editor. The ontology comprises descriptions of the
FHW exhibits in OWL, as well as linguistic and user type
information in RDF. The system supports both English and
Greek.
Although an evaluation of the robotic guide has not been
performed yet in real conditions, an initial version of the
dialogue system has already been examined by FHW. In
particular, the dialogue model, the richness of the generated text by the NLG, and the ontology are currently under
revision based on user interaction data.
In the future we envisage the integration of an affective
unit into the dialogue system. There is considerable research that suggests that user emotion recognition and robotic emotion expression may lead to more natural forms
of communication (André et al., 2004).
Acknowledgements
This work was carried out in the context of project XENIOS, which was co-funded by the Greek General Secretariat and the European Union.
7.
References
E. André, L. Dybkjær, W. Minker, and P. Heisterkamp, editors. 2004. Affective Dialogue Systems. Springer.
I. Androutsopoulos, J. Oberlander, and V. Karkaletsis.
2007. Source authoring for multilingual generation of
personalised object descriptions. Natural Language Engineering, 13(3):191–233.
M. Bennewitz, F. Faber, M. Schreiber, and S. Behnke.
2005. Towards a humanoid museum guide robot that interacts with multiple persons. In Proceedings of the 5th
IEEE-RAS International Conference on Humanoid Robots.
D. Bilidas, M. Theologou, and V. Karkaletsis. 2007.
Enriching OWL Ontologies with Linguistic and Userrelated Annotations: the ELEON system. In Proceeding
of International Conference on Tools with Artificial Intelligence (ICTAI). IEEE Computer Society Press.
K. Bontcheva and Y. Wilks. 2004. Automatic report generation from ontologies: the MIAKT approach. In Proceedings of the 9th International Conference on Applications
59
of Natural Language to Information Systems, pages 324–
335, Manchester, UK.
K. Bontcheva, V. Tablan, D. Maynard, and H. Cunningham. 2004. Evolving GATE to meet new challenges in
language engineering. Natural Language Engineering,
10(3/4):349–373.
C. Chiu. 2004. The Bryn Mawr Tour Guide Robot. Ph.D.
thesis, Bryn Mawr College.
D. Galanis and I. Androutsopoulos. 2007. Generating multilingual descriptions from linguistically annotated OWL
ontologies: the NATURAL OWL system. In Proceedings
of the 11th European Workshop on Natural Language
Generation, Schloss Dagstuhl, Germany.
A. Isard, J. Oberlander, I. Androutsopoulos, and C. Matheson. 2003. Speaking the users’ languages. IEEE Intelligent Systems, 18(1):40–45.
M. F. McTear. 2004. Spoken Dialogue Technology. Toward
the Conversational User Interface. Springer.
A. Melengoglou. 2002. Multilingual aggregation in the
M - PIRO system. Master’s thesis, School of Informatics,
University of Edinburgh, UK.
D. Milward and M. Beveridge. 2003. Ontology-based dialogue systems. In 3rd Workshop on Knowledge and Reasoning in Practical Dialog Systems.
M. O’Donnell, C. Mellish, J. Oberlander, and A. Knott.
2001. ILEX: an architecture for a dynamic hypertext generation system. Natural Language Engineering,
7(3):225–250.
G. Paliouras, A. Mouzakidis, C. Ntoutsis, A. Alexopoulos,
and C. Skourlas. 2006. PNS: Personalized Multi-Source
News Delivery. In Proceedings of the 10th International
Conference on Knowledge-Based & Intelligent Information & Engineering Systems (KES), UK.
E. Reiter and R. Dale. 2000. Building natural language
generation systems. Cambridge University Press.
S. Thrun, M. Bennewitz, W. Burgard, A.B. Cremers,
F. Dellaert, D. Fox, D. Haehnel, C. Rosenberg, N. Roy,
J. Schulte, and D. Schulz. 1999. MINERVA: A second
generation mobile tour-guide robot. In Proceedings of
the IEEE International Conference on Robotics and Automation (ICRA).
G. Xydas and G. Kouroupetroglou. 2001. The DEMOSTHENES Speech Composer. In Proceedings of the
4th ISCA Tutorial and Workshop on Speech Synthesis,
Perthshire, Scotland.
60
A Semantic Wiki Approach to Cultural Heritage Data Management
René Witte, Thomas Gitzinger, Thomas Kappler, Ralf Krestel
Institut für Programmstrukturen und Datenorganisation (IPD)
Universität Karlsruhe (TH), Germany
Abstract
Providing access to cultural heritage data beyond book digitization and information retrieval projects is important for delivering advanced
semantic support to end users, in order to address their specific needs. We introduce a separation of concerns for heritage data
management by explicitly defining different user groups and analyzing their particular requirements. Based on this analysis, we developed
a comprehensive system architecture for accessing, annotating, and querying textual historic data. Novel features are the deployment of a
Wiki user interface, natural language processing services for end users, metadata generation in OWL ontology format, SPARQL queries on
textual data, and the integration of external clients through Web Services. We illustrate these ideas with the management of a historic
encyclopedia of architecture.
1.
Introduction
contemporary work. In our example scenario, the handbook on architecture, these are today’s architects that need
information on the building processes and materials used,
e.g., within a restoration project of an old building. Here,
the historic material contains knowledge that is not readily
accessible in modern sources. Another example for such a
user group are musicians dealing with old music scores and
their descriptions, or lexicographers analyzing documents
for the development of dictionary entries.
(3) Laypersons: Historic materials are a fascinating
source of knowledge, as they preserve information over
centuries. Providing widespread online access to materials
that are otherwise only available in a controlled environment
to scientists due to their fragile nature is perhaps one of the
greatest benefits of digitization projects.
(4) Computational Linguists: Similarly to practitioners,
linguists are often interested in historic documents from a
functional point of view. However, their domain focuses
on the properties of the language and its development over
time rather than the underlying domain of discourse. They
also have particular requirements for corpus construction,
access, and annotation to support automated NLP analysis
workflows.
The amount of publicly available knowledge increases faster
than we can imagine—hence the term “Information Explosion” used by several authors (Lyman and Varian, 2003;
Sweeney, 2001). With the barrage of newly created content—
news, blogs, web pages, research papers—existing, “analog”
documents and their users often receive less attention than
the quality of the content deserves.
In this paper, we present the results from a project aimed at
developing enhanced semantic support for users of textual
cultural heritage data. A particular feature of our approach is
the integration of different concerns into a single, cohesive
system architecture that addresses requirements from end
users, software engineering aspects, and knowledge discovery paradigms. The ideas were implemented and tested with
a historic encyclopedia of architecture and a number of different user groups, including building historians, architects,
and NLP system developers.
2.
User Groups and Requirements
Nowadays, the baseline for cultural heritage data management of book-type publications is the production of
a scanned (digitized) version that can be viewed and distributed online, typically with some kind of Web interface.
Before we can deliver more advanced access methods, we
have to be more precise about the targeted end users. Who
needs access to heritage data, and for what purpose?
2.1.
2.2. Detected Requirements
We can now derive a number of requirements a system needs
to fulfill, based on the user groups defined above:
Web Interface. To make the historic data available over
the Internet, and to provide easy access within a familiar
metaphor, the system needs to support a Web interface. This
concerns all user groups to various degrees, but in particular
the historians and laypersons.
Annotation Support. Users working with the historic
data from a scientific point of view—in particular group
(1)—often need to comment, add, and collaborate on the
historic data. This should be supported within the same
interface as the primary (historic) data, to avoid unnecessary
context and application switches for the end users. At the
same time, these annotations must be maintained by the
architecture on clearly separated layers, to keep the integrity
of the historic data intact.
Corpus Generation. While a Web interface is helpful for
a human user, automated analyses using NLP tools and
User Groups
Within our project, we had to consider the requirements from
four different user groups; each of them having a different
background and expectations concerning the management
of historic textual data.
(1) Historians: Within this group, we target users that deal
with historic material from a scientific motivation, namely,
historians. They require an electronic presentation that provides for a direct mapping to the printed original, e.g., for citation purposes. Additionally, semantic analysis tools should
support their work through the formulation and verification
of hypotheses.
(2) Practitioners: Under this group, we are concerned
with users that need access to the historic material for their
61
frameworks (user group (4)) can be better supported with
a corpus in a standard (XML-based) markup, since HTML
pages generated through Web frameworks typically mix content and layout information (menus, navigation bars, etc.).
Thus, the architecture should provide a separate corpus that
is automatically derived from the historic data and contains
appropriate markup (for headlines, footnotes, figure captions, etc.). Ideally, it should allow to cross-link entities
with the Web interface.
As outlined in (Mavrikas et al., 2004), access to cultural
heritage data available in natural language can be facilitated
using various NLP techniques. In the context of the Semantic Web, the proposed system extracts CH data from
different sources in the Internet and processes the data afterwards. An ontology (Doerr, 2003) is used to organize
the mined data. Templates are used to extract relevant information, and the use of multi-document summarization
is also proposed, as a way to present relevant information
in a condensed way to the user. Here, we present an actual
implementation of a system addressing these problems and
extend the use of ontologies to allow easy browsing and
querying of the document content for different user groups.
Another approach based on the CIDOC-CRM1 ontology is
presented in (Généreux, 2007). The system described there
consists of two parts, one for extracting CH knowledge
from natural language texts and saving the information in
the ontology format, and one for using natural language to
query the database. The natural language is reformatted to a
SPARQL query using WordNet. This approach, in contrast
to our system, stresses more the search aspect to find relevant
data and offers no further possibilities for collaboration or
processing of the data.
In (Sinclair et al., 2005), a system is presented that enables
the user to explore, navigate, link, and annotate digitized
cultural heritage artifacts like videos, photos, or documents.
The system also supports user-generated descriptions and
content. The focus in this project lies on the integration of
the different metadata formats of the source content, whereas
we additionally focus on the processing and collaboration
part.
From a technical perspective, semantic extensions to Wiki
systems based on Semantic Web technologies like OWL ontologies and RDF are similar in that they provide the means
for content structuring beyond the syntactical level. In these
systems, the properties of and relations between objects can
be made explicit, with the Wiki system “knowing” about
them. This allows for automated processing of Wiki content,
e.g., through software agents. Current implementations of
these ideas can be found in systems like Semantic MediaWiki (SMW) (Krötzsch et al., 2006) or IkeWiki (Schaffert,
2006). It is important to note that these tools are different
from and complementary to our approach: While in our
context, the content of a Wiki is subject to semantic analysis
via NLP methods (with the Wiki engine itself not needing
to have semantic capabilities), semantic Wikis like SMW
have explicit notational and internal semantic capabilities.
Using a semantic Wiki in our system in the future would
allow the Wiki engine itself direct access to the facts derived
from semantic text analysis.
NLP Services. For large collections of (historic) documents, manual inspection of all content or even a subset
obtained through information retrieval (IR) is not feasible.
Here, NLP analyses can deliver additional benefit to end
users, in particular groups (1)–(3), by integrating NLP analysis services (and their results) into the overall architecture.
It should allow the execution of any service, developed by
user group (4), and also deliver the results back to the clients.
Examples for such NLP services are summarization, index
generation, or named entity detection.
Metadata Generation. While NLP results can be useful
for a human user, we also need to support further automated
analysis workflows. User group (2) in particular requires
access to the historic data, as well as its metadata, from
external tools and applications relevant for their domain. To
support external access to metadata from many different
clients, the architecture should be capable of generating
standards-compliant data formats, such as RDF and OWL.
Application Integration. As pointed out in the last requirement, external applications should be provided with
automated access to the historic data and its metadata. Generally speaking, this requires the introduction of a client/server model, where the communication, like the metadata
format, should use open, established standards.
3.
Related Work
Before we describe our approach in detail, we discuss related
work relevant for the detected requirements.
The Cultural Heritage Language Technologies (CHLT)
project (Rydberg-Cox, 2002; Rydberg-Cox, 2005) describes
the use of NLP methods to help students and scholars to
work with classic Greek and Latin corpora. Similar to our
approach, collaboration is an important goal of the project.
Not only for sharing metadata about the text itself, but also
to offer users the possibility to annotate, comment, or correct the results of automated analysis. This metadata can
also contain hyperlinks to connect related texts with each
other. The importance of correct morphological analysis is
stressed as a baseline technology for users in the humanities,
a statement which is also reflected in our work by integrating
a self-learning lemmatizer for the German language (Perera
and Witte, 2005) for accurate index generation. Further processing in the CHLT project includes information retrieval
and data visualization. Identifying keywords, clustering subsets of the data, and visualizing the resulting groups supports
the users in grasping concepts or performing search. In contrast, our approach uses open, standardized data formats like
an automatically populated ontology to facilitate searching
and browsing through the corpus and a Wiki system to share
information between users.
4.
Semantic Heritage Data Management
In this section, we present our approach to cultural heritage
data management, which integrates a number of different
technologies in order to satisfy the requirements of the various user groups: (i) A Wiki user interface, (ii) text mining
support using an NLP framework, (iii) Semantic Web ontologies based on OWL and RDF for metadata management,
1
62
CIDOC Conceptual Reference Model, http://cidoc.ics.forth.gr/
Tier 1: Clients
Tier 2: Presentation and Interaction
Tier 3: Analysis and Retrieval
Tier 4: Resources
NLP Subsystem
Service Information
Language Services
Language
Service
Descriptions
Service Invocation
Information Extraction
Web Server
Desktop
Client
Client Side Abstraction Layer
Plugin
OpenOffice.org
Writer
NLP Service Connector
NLP/Wiki Connector
Automatic Summarization
Question Answering
Documents
Index Generation
Web
Client
Wiki Engine
Information Retrieval
Document
Metadata
Figure 1: System architecture overview
and (iv) W3C Web Services for application integration. We
first present an overview of our system in the next subsection. The various subsystems are illustrated using examples
from a productive, freely accessible2 Web resource built
around the German Handbuch der Architektur (handbook
on architecture) from the 19th century, described in detail
in Section 4.2. The digitization process is described in Section 4.3. Necessary format conversions for the digital version are covered in Section 4.4. To support our user groups,
we integrated several NLP analysis services, which are covered in Section 4.5. Finally, our semantic extensions for
generating OWL/RDF metadata and application integration
are covered in Section 4.6.
“Client-Side Abstraction Layer” (CSAL) facilitates connecting clients by providing common communication and data
converting functionality.
The clients communicate with a Web server on Tier 2, behind which we find the Wiki engine and a software module
labeled “NLP Service Connector.” The functionality of this
module is offered as an XML Web service, as standardized
by the W3C.4 This means that there is a publicly accessible
interface definition, written in the Web Service Description
Language (WSDL), from which clients know how to use
the offered functionality. The functionality itself is used
through a Web service endpoint, to which the client sends
and from where it receives messages. The main task of
the NLP Service Connector is to receive input documents
and have the NLP subsystem (Tier 3) perform various text
analysis procedures on them. A sub-module of the NLP
Service Connector, labeled “NLP/Wiki Connector,” allows
for the automatic retrieval, creation, and modification of
Wiki content.
Finally, on Tier 4, we have metadata on the employed text
analysis services (top), which the NLP Service Connector
requires in order to operate these services. The bottom
rectangle contains the documents maintained by the Wiki
system as well as their metadata, which might have been
provided by hand, or generated through automatic analysis
methods.
4.1. Architectural Overview
As stated above, our goal is the development of a unified architecture that fulfills the requirements (Section 2.2.) of the
different user groups defined in Section 2.1., by integrating
means for content access, analysis, and annotation.
One of the central pieces of our architecture is the introduction of a Wiki system (Leuf and Cunningham, 2001). Wiki
systems provide the Web interface stipulated in our first
requirement, while also allowing users to add meta-content
in form of separate discussion or annotation pages. This
capability directly addresses our second requirement, by
allowing users to discuss and collaborate on heritage data,
using an online tool and a single interface, while keeping
the original data intact.3
Other clients, NLP services, and the actual content have to
be integrated into this model. Figure 1 shows how these and
the remaining components are systematically assembled to
form the overall architecture of our system.
The architecture comprises four tiers. Tier 1 consists of
clients that the users employ to access the system. Plug-in
capable existing clients, like the OpenOffice.org application
suite, can also be extended to be integrated with our architecture. New applications can have that functionality built
in, like the “Desktop Client” depicted in the diagram. The
4.2. Source Material
We implemented and evaluated the ideas described here for a
particular set of historic documents: the German Handbuch
der Architektur, a comprehensive multi-volume encyclopedia of architecture.5 The full encyclopedia was written
between the late 19th and early 20th century; It aimed to
include all architectural knowledge at the time, both past
and present, within the fields of architectural history, architectural styles, construction, statics, building equipment,
physics, design, building conception, and town planning.
The full encyclopedia comprises more than 140 individual
2
4
See http://durm.semanticsoftware.info
Assuming the Wiki has been properly configured for this scenario; the technical details depend on the concrete Wiki system.
Web Services Architecture, http://www.w3.org/TR/ws-arch/
Edited by Joseph Durm (b14.2.1837 Karlsruhe, Germany,
d3.4.1919 ibidem) and three other architects since 1881.
3
5
63
Figure 2: Source material examples: Scanned pages from Handbuch der Architektur (1900)
markup.
publications and contains at least 25 000 pages.
Due to the ambitious scope, the long publication process,
and the limitations of the technologies available at that time,
it is extremely difficult to gain an overview of a single topic.
Information is typically distributed over several parts containing a number of volumes, which in turn are split into
books. Most of these do not contain any kind of index. In
addition, some of the volumes were edited and reprinted and
a supplement part was added.
Due to funding limitations, we only dealt with a single
volume6 within the project described in this paper. However,
the concepts and technologies have been developed with the
complete dataset in mind.
4.4.1. TUSTEP Format
TUSTEP is a toolkit for the “scientific work with textual
data” (Uni, 2008), consisting of a document markup standard along with tools for text processing operations on
TUSTEP documents. The markup is completely focused on
layout, so that the visual structure of printed documents can
be captured well. Structurally, it consists both of XML-like
elements with an opening and closing tag, such as <Z> and
</Z> for centered passages; and elements serving as control
statements, such as #H: for starting text in superscript. The
control statements remain in effect until another markup
element cancels them out, such as #G: for adjusting the
following text on the baseline.
TUSTEP predates XML, and while it is still in use at many
universities, we found it makes automatic processing difficult. The control statements, for instance, make it hard
to determine the range of text they affect, because their
effect can be canceled by different elements. In addition,
in the manual digitization process, markup was applied inconsistently. Therefore, we chose to first convert the data
to a custom XML format, designed to closely match the
given TUSTEP markup. This also enabled easier structural
analysis and transformation of the text due to the uniform
tree structure of XML and the availability of high-quality
libraries for XML processing.
4.3. Digitization and Error Correction
The source material is first digitized using specialized book
scanners, producing a TIFF file for each physical page; in
our case, with a grayscale resolution of 600dpi.
In a second step, the image files need to be converted to
machine-readable text to support, amongst others, NLP analysis and metadata generation. We initially planned to automate this process using OCR software. However, due to
the complex layout of the original material (see Figure 2),
which contains an abundance of figures, graphs, photos, tables, diagrams, formulas, sketches, footnotes, margin notes,
and mixed font sizes, as well as the varying quality of the
100-year old source material, this proved to be too unreliable.
As the focus of this project was on developing enhanced
semantic support for end users, not basic OCR research,
we decided to manually convert the source material into an
electronic document. This provided for not only a faster
and more reliable conversion, but also accurately captured
layout formation in explicit markup, such as footnotes, chapter titles, figure captions, and margin notes. This task was
outsourced to a Chinese company for cost reasons; Manual conversion was performed twice to allow an automatic
cross-check for error detection. The final, merged version
contained only a very small amount of errors, which were
eventually hand-corrected during the project.
4.4.2. Custom XML
We developed a custom tool to transform TUSTEP data into
XML. The generated XML data intends to be as semantically
close to the original markup as possible; as such, it contains
mostly layout information such as line and page breaks and
font changes. Except for the exact placement of figures
and tables, all such information from the original book is
retained.
Parsing the XML into a DOM8 representation provides for
easy and flexible data transformation. The resulting XML
format can be directly used for NLP corpus generation.
4.4.3. Wiki Markup
To make the historic data accessible via a Wiki, we have to
further transform it into the data format used by a concrete
Wiki engine. Since we were dealing with an encyclopedic
original, we chose the MediaWiki9 system, which is best
known for its use within the Wikipedia10 projects.
4.4. Format Transformation and Wiki Upload
The digitized content was delivered in the TUSTEP7 format.
This content was first converted to XML, and finally to Wiki
6
E. Marx: Wände und Wandöffnungen (Walls and Wall Openings). In “Handbuch der Architektur,” Part III, Volume 2, Number I,
Second edition, Stuttgart, Germany, 1900. Contains 506 pages with
956 figures.
7
TUebingen System of TExt processing Programs (TUSTEP),
http://www.zdv.uni-tuebingen.de/tustep/tustep eng.html
8
Document Object Model (DOM), http://www.w3.org/DOM/
MediaWiki, http://en.wikipedia.org/wiki/MediaWiki
10
Wikipedia, http://www.wikipedia.org
9
64
A challenging question was how to perform the concrete
conversion from content presented in physical book layout
to Wiki pages. Obviously, translating a single book page
does not translate well into a single web page. We first
attempted to translate each book chapter into a single page
(with its topic as the Wiki entry). However, with only 15
chapters in a 500-page book, the resulting Web pages were
too long to be used comfortably in the MediaWiki interface.
Together with our end users, we finally decided to convert
each sub-chapter (section) into a single Wiki page, with
additional internal structuring derived from the margin notes
preserved by the manual conversion.
MediaWiki uses the markup language Wikitext, which was
designed as a “simplified alternative to HTML,”11 and as
such offers both semantic markup, like headings with different levels, as well as visual markup, like italic or bold text.
Its expressiveness is largely equal to that of HTML, despite
the simplified approach, because it lets users insert HTML
if Wikitext does not suffice.
Figure 3: The Wiki interface integrating digitized text,
scanned originals, and separate “Discussion” pages
(see Figure 1, Tier 3, “NLP Subsystem”) by GATE, the
General Architecture for Text Engineering (Cunningham et
al., 2002). To allow a dynamic discovery of newly deployed
language services, we added service descriptions written in
OWL to our architecture (see Section 4.1.).
Language services should help the users to find, understand,
relate, share, and analyze the stored historic documents. In
the following subsections, we describe some of the services
we deployed in our implementation to support users of the
historic encyclopedia, including index generation, automatic
summarization, and OWL metadata generation.
Example: Footnote conversion. Footnotes were delivered in TUSTEP in the form #H:n#G:) for each footnote
n. The markup indicates text being set to superscript (#H:),
then back to the standard baseline (#G:). The footnote
reference in the text and the anchor in the footnote section
of a page have the same markup, as they look the same. The
tool converting to XML locates footnotes using a regular
expression, and creates <footnote to="n" /> resp.
<footnote from="n">...</footnote> tags. Finally,
the conversion to Wikitext transforms the references to
<span id="fn8ref" /><sup>[[#fn8|8)]]</sup>.
The HTML sup tag sets the text as superscript, and its content is a link to the anchor “fn8” on the same page, with the
link text simply being “8”. The footnote itself is represented
by <span id="fn8"/>’’8)’’ ... [[#fn8ref|ˆ]].
We see the anchor linked to from the reference, and vice
versa a link to jump back upwards to the reference.
4.5.1. Index Generation
Many documents—like the discussed architectural
encyclopedia—do not come with a classical back-of-thebook index. Of course, in the absence of an index, full-text
search can help to locate the various occurrences of a
single term, but only if the user already knows what he is
looking for. An index listing all nouns with their modifiers
(adjectives), with links to their locations of occurrence, can
help the user find useful information he was not expecting,
which is especially important for historical documents,
which often contain terminology no longer in use.
For our index, we process all noun phrases found in the analyzed texts. For each noun phrase, we compute the lemma of
the head noun and keep track of its modifiers, page number,
and corresponding Wiki page. To deal with the problem of
correctly lemmatizing historic terminology no longer in use,
we developed a self-learning lemmatizer for German (Perera and Witte, 2005). Nouns that have the same lemma are
merged together with all their information. Then, we create
an inverted index with the lemma as the main column and
their modifiers as sub-indexes, as shown in Figure 4. The
generated index is then uploaded from the NLP subsystem
into the Wiki through a connector (“NLP/Wiki Connector”
in Figure 1).
4.4.4. Wiki Interface Features.
The conversion to Wikitext inserts further information for
the Wiki users, such as links to scans of the original pages,
and link/anchor combinations to emulate the page-based
navigation of the book (see Figure 3). For instance, the
beginning of page 211, which is indicated in TUSTEP by
@@[email protected]<S211><, looks as follows in the resulting Wikitext:
<span id="page10" />
’’’Seite 211 ([[Media:S211_large.gif|Scan]])’’’
[[Image:S211_large.gif|thumb|200px|Scan der
Originalseite 211]]
4.5.
NLP Integration
One of the main goals of our work is to support the end
users—groups (1) to (3)—with semantic analysis tools based
on NLP. To make our architecture independent from the
application domain (architecture, biology, music, . . .) and
their custom NLP analysis pipelines, we developed a general
integration framework that allows us to deploy any kind of
language service. The management, parametrization, and
execution of these NLP services is handled in our framework
11
4.5.2. Automatic Summarization
Large text corpora make it impossible for single users to
deal with the whole documents in total. The sheer amount
of information encoded in natural language in huge text
collections poses a non-trivial challenge to information sys-
Wikitext, http://en.wikipedia.org/wiki/Wikitext
65
Figure 4: NLP-generated full text index, integrated into the
Wiki interface (page numbers are hyperlinks to Wiki pages)
tems in order to adequately support the user. To find certain
information, to get an overview of a document, or just to
browse a text collection, automatic summarization (Mani,
2001) offers various methods of condensing texts.12
Short, headline-like summaries (around 10 words) that incorporate the most important concepts of a document or a
Wiki page facilitate the search for particular information
by giving a user an overview of the content at a glance. In
addition, full-text summaries can be created for each page,
e.g., with a length of 100 words or more. These summaries
in free-text form can be read much more quickly than a
full-length article, thereby helping a user to decide which
Wiki pages he wants to read in full.
More advanced types of summaries can support users during
both content creation and analysis. Multi-document summaries can combine knowledge from several pages within a
Wiki or even across Wiki systems. Update summaries keep
track of a user’s reading history and only present information
he has not read before, thereby further reducing the problem
of information overload. And focused summaries enable the
user to formulate a query (natural language questions) the
generated summary focuses on. This is especially useful to
get a first impression of the available information about a
certain topic in a collection. In (Witte et al., 2005), we illustrate the usefulness of focused summaries for a particular
architectural scenario.
Figure 5: An ontology instance created through NLP
indices, and question-answering all produce new natural
language texts, which are convenient for humans. But they
are less useful for providing further automated access to
the historic data, e.g., through desktop tools targeted at user
group (2). In our example scenario, the architects need to integrate the historic knowledge “stored” in the encyclopedia
within contemporary architectural design tools: While viewing a certain construction element, the relevant content from
the handbook should be extracted and presented alongside
other project information. This requires the generation of
metadata in a machine-processable data format. In our architecture, this is provided through the NLP-driven population
of formal ontologies. We discuss our ontology model in the
next subsection, followed by a description of the automatic
population process and the querying of the result format.
4.6.1. Ontology Model
Our NLP-generated metadata is formally represented using
the Web Ontology Language (OWL),13 which is a standard
defined by the World Wide Web Consortium (W3C). Specifically, we use the sub-format OWL-DL, which is based on
description logics (DL). OWL is also the foundation of the
Semantic Web initiative, which allows us to immediately
make use of a large variety of tools and resources developed
for OWL-based information processing (editors, storage systems, query languages, reasoners, visualization tools, etc.).
Our ontology has two parts: a document ontology describing
the domain of NLP (documents, sentences, NPs, coreference
chains, etc.) and a domain ontology. While the document
ontology is independent of the content in the historic documents, the domain ontology has to be developed specifically
for their discourse domain. In our example, this ontology
needs to contain architectural concepts, such as doors, walls,
or windows. By combining both ontologies, we can run
semantic queries against the ontology, e.g., asking for all
sentences where a certain concept appears. The incorporation of CIDOC/CRM could extend our model in the future.
4.5.3. Other NLP Services
The examples presented so far are by no means exhaustive.
Depending on the type of data under investigation and the demands of the users concerned with their analysis (groups (1)
and (2)), additional NLP services will need to be introduced.
Due to our service-oriented approach (cf. Section 4.1.), new
services can be added at any time, as they are automatically
detected by all connected clients through the metadata repository, without any changes on the client side. Likewise, new
user clients can be added dynamically to the architecture,
without requiring any changes to the NLP server.
4.6. Semantic Extensions
The NLP analysis services introduced so far are aimed at
supporting the user groups (1) and (3): Summaries, full-text
Document Ontology Model. Our document ontology
models a number of concepts relevant for the domain of
12
See, e.g., the Document Understanding Conference (DUC),
http://duc.nist.gov
13
66
OWL, http://www.w3.org/2004/OWL/
NLP. One of the main concepts is document, representing
an individual text processed by an NLP pipeline, containing:
the title of the document; its source address (typically a
URL or URI); and a relation containsSentence between a
document and all its sentences.
Likewise, sentences are also represented by an ontology
class, with: the start and end position (beginLocation, endLocation) within the document, given as character offset;
the sentence’s content, stored as plain text, i.e., without additional markup; and a relation contains between a sentence
and all named entities that have been detected in it.
Each of the named entities has, in addition to its ontology class, a number of additional properties: a unique id
(idPropOf ) generated for this instance; the page number
(originalPageNumber), where the instance can be found in
the (printed) source; and the full URL (pageURL) for direct
access to the instance in the Wiki system.
Additionally, we can represent the result of the coreference resolution algorithm using the OWL language feature
sameAs: If two instances appear in the same coreference
chain, two separate ontology instances are created (containing different ids and possibly different page/URL numbers),
but both instances are included in such a sameAs relation.
This allows ontology reasoners to interpret the syntactically
different instances as semantically equivalent. Additionally,
a relation corefStringWithId is created for every entity in the
coreference chain, refering to its unique id stored in the idPropOf property; and the content of the sentence containing
the co-refering entity is stored in corefSentenceWithId.
NLP Preprocessing:
Tokenisation, Sentence Splitting, ...
Domain
Ontology
Part−of−Speech Tagging
Noun Phrase Chunking
Document
Ontology
Lemmatisation
Onto−Gazetteer: Assign
Assign Ontology
Combined
Ontology Classes to Tokens
Classes to Tokens
Ontology
Grammars for
Named Entity Detection
Ontology−Aware
Grammar Rules
Ontology
Population
Coreference Resolution
Populated
Ontology
OWL Ontology Export
Figure 6: NLP pipeline for ontology population
Finally, the created instances are exported into the result
ontology, combining a number of domain and document features. An example instance, of the ontology class Kalkmörtel
(lime mortar), is shown in Figure 5.
4.6.3. Ontology Queries
The automatically populated ontology represents a machinereadable metadata format that can be queried through a
number of standardized ontology query languages, such as
SPARQL.15 Queries are a much more expressive paradigm
for analyzing text mining results than simple IR; in particular, if a domain model is available, they allow queries over
the analyzed documents on a semantic level.
An example SPARQL query is shown in Figure 7. The
query shown in the left box represents the question “Which
building materials are mentioned in the handbook together
with the concept ‘Mauer’ (wall), and on which page?” The
result of this query (executed using Protégé16 ), is shown
on the right. The first column (“type”) shows what kind
of entity (stone, plaster, concrete, . . .) was found, i.e., a
sub-class of “material” in the domain ontology. The results
can now be directly inspected by the user or used for further
automatic processing by another application.
More abstractly speaking, ontology queries support automated problem-solving using a knowledge base. A user of
our system, like a historian, might want to formulate hypotheses concerning the source material. Translated into an
OWL query, the result can be used to confirm or refute the
hypothesis. And as a standardized NLP result format, it also
facilitates direct integration into an end-user application or
a larger automated knowledge discovery workflow.
Domain Ontology Model. In addition to the generic NLP
ontology, a domain-specific ontology can be plugged into
the system to allow further structuring of the NLP results. If
such an ontology is developed, it can also be used to further
facilitate named entity detection as described below.
In our approach, we rely on a hand-constructed ontology of
the domain. This could be enhanced with (semi-)automatic
ontology enrichment or ontology learning. In general, the
design of the domain ontology needs to take the requirements of the downstream applications using the populated
ontology into account.
4.6.2. Automatic Ontology Population
We developed an ontology population NLP pipeline to automatically create OWL instances (individuals, see Figure 5)
for the ontology described above. An overview of the workflow is shown in Figure 6.
The pipeline runs on the XML-based corpus described in
Section 4.4. After a number of standard preprocessing steps,
including tokenization, POS tagging, and NP chunking,
named entities (NEs) are detected using a two-step process. First, an OntoGazetteer (Bontcheva et al., 2004) labels
each token in the text with all ontology classes it can belong
to. And secondly, ontology-aware grammar rules written in
the JAPE14 language are used to find named entities (NEs).
Evaluation of the correctness of the generated instances can
be conducted using precision and recall measures (Maynard
et al., 2006).
4.6.4. Application Integration
The populated ontology also serves as the basis for our final
requirement, application integration. With “application” we
15
14
Java Annotations Pattern Engine, a regular expression-based
language for writing grammars over document annotation graphs.
16
67
SPARQL, http://www.w3.org/TR/rdf-sparql-query/
Protégé, http://protege.stanford.edu/
Figure 7: Posing a question to the historic knowledge base through a SPARQL query against the NLP-populated ontology
mean any end-user accessible system that wants to integrate
the historic data within a different context. For example, in
a museum setting, such an application might allow a visitor
to access content directly relevant to an artifact. A lexicographer might want to query, navigate, and read content
from historical documents while developing a lexical entry.
And in our application example, an architect needs access
to the knowledge stored in the handbook while planning a
particular building restoration task. Here, construction elements displayed in a design tool (such as window or window
sill) can be directly connected with the ontological entities
contained in the NLP-populated knowledge. This allows
an architect to view relevant content down to the level of
an individual construction element using the named entities,
while retaining the option to visit the full text through the
provided Wiki link.
5.
Martin Doerr. 2003. The CIDOC Conceptual Reference Module: An Ontological Approach to Semantic Interoperability of
Metadata. AI Mag., 24(3):75–92.
Michel Généreux. 2007. Cultural Heritage Digital Resources:
From Extraction to Querying. In Proceedings of the Workshop
on Language Technology for Cultural Heritage Data (LaTeCH
2007), pages 41–48, Prague, Czech Republic, June. ACL.
Markus Krötzsch, Denny Vrandečić, and Max Völkel. 2006. Semantic MediaWiki. In Isabel Cruz, Stefan Decker, Dean Allemang, Chris Preist, Daniel Schwabe, Peter Mika, Mike Uschold,
and Lora Aroyo, editors, The Semantic Web – ISWC 2006, volume 4273 of LNCS, pages 935–942. Springer.
Bo Leuf and Ward Cunningham. 2001. The Wiki Way, Quick
Collaboration on the Web. Addison-Wesley.
Peter Lyman and Hal R. Varian. 2003. How Much Information?
I. Mani. 2001. Automatic Summarization. John Benjamins B.V.
Efthimios C. Mavrikas, Nicolas Nicoloyannis, and Evangelia
Kavakli. 2004. Cultural Heritage Information on the Semantic Web. In Enrico Motta, Nigel Shadbolt, Arthur Stutt, and
Nicholas Gibbins, editors, EKAW, volume 3257 of Lecture Notes
in Computer Science, pages 477–478. Springer.
D. Maynard, W. Peters, and Y. Li. 2006. Metrics for Evaluation of
Ontology-based Information Extraction. In Proceedings of the
4th International Workshop on Evaluation of Ontologies on the
Web (EON 2006), Edingburgh, UK, May.
Praharshana Perera and René Witte. 2005. A Self-Learning
Context-Aware Lemmatizer for German. In Proc. of Human
Language Technology Conference and Conference on Empirical
Methods in Natural Language Processing (HLT/EMNLP 2005),
pages 636–643, Vancouver, BC, Canada, October 6–8.
Jeffrey A. Rydberg-Cox. 2002. Cultural Heritage Language Technologies: Building an Infrastructure for Collaborative Digital
Libraries in the Humanities. Ariadne, 34, December.
Jeffrey A. Rydberg-Cox. 2005. The Cultural Heritage Language
Technologies Consortium. D-Lib Magazine, 11(5), May.
Sebastian Schaffert. 2006. IkeWiki: A Semantic Wiki for Collaborative Knowledge Management. In WETICE, pages 388–396.
Patrick Sinclair, Paul Lewis, Kirk Martinez, Matthew Addis,
Adrian Pillinger, and Daniel Prideaux. 2005. eCHASE: Exploiting Cultural Heritage using the Semantic Web. In 4th International Semantic Web Conference (ISWC 2005), Galway,
Ireland, November 6–10.
L. Sweeney. 2001. Information Explosion. In L. Zayatz, P. Doyle,
J Theeuwes, and J. Lane, editors, Confidentiality, Disclosure,
and Data Access: Theory and Practical Applications for Statistical Agencies. Urban Institute, Washington, DC.
Universität Tübingen – Zentrum für Datenverarbeitung, 2008.
TUSTEP: Handbuch und Referenz. Version 2008.
René Witte, Petra Gerlach, Markus Joachim, Thomas Kappler, Ralf
Krestel, and Praharshana Perera. 2005. Engineering a Semantic
Desktop for Building Historians and Architects. In Proc. of the
Semantic Desktop Workshop at the ISWC 2005, volume 175 of
CEUR, pages 138–152, Galway, Ireland, November 6.
Summary and Conclusions
To support users in the cultural heritage domain, a precise
analysis of the different user groups and their particular requirements is essential. In this paper, we present a holistic
approach based on a unified system architecture that highlights the many inter-dependencies in supporting different
groups with particular features, aimed at different use cases:
Historians have the support of NLP analysis tools and a
user-friendly Web-based access and collaboration tool build
around a standard Wiki system. Laypersons also benefit
from these user-friendly features, while practitioners—in
our scenario building architects—can additionally use NLPgenerated ontology metadata for direct application integration. Finally, our approach also supports computational
linguists through corpus construction and querying tools.
The experience from the implemented system using the
example of a historical encyclopedia of architecture demonstrates the usefulness of these ideas. Finally, providing a
machine-readable knowledge base that integrates textual instances and domain-specific entities is consistent with the
vision of the Semantic Web, which has the potential to further enhance knowledge discovery for cultural heritage data.
6.
References
Kalina Bontcheva, Valentin Tablan, Diana Maynard, and Hamish
Cunningham. 2004. Evolving GATE to Meet New Challenges
in Language Engineering. Natural Language Engineering.
H. Cunningham, D. Maynard, K. Bontcheva, and V. Tablan. 2002.
GATE: A framework and graphical development environment
for robust NLP tools and applications. In Proc. of the 40th
Anniversary Meeting of the ACL. http://gate.ac.uk.
68
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement