Verbs in specialised corpora : from manual corpus-based

Verbs in specialised corpora : from manual corpus-based
Verbs in specialised corpora: from manual corpus-based description to automatic
extraction in an English-French parallel corpus
Natalie Kübler, CIEL, Université Paris 7–Denis Diderot, 2, place Jussieu,
F-75251 Paris Cedex 05, [email protected]
Cécile Frérot, ERSS, UMR 5610, Université de Toulouse-Le Mirail,
5, allées A. Machado, F-31058 Toulouse Cedex, [email protected]
This paper tackles the issue of verbs in specialised corpora in the view of term extraction. Corpus-based manual
descriptions to be used in various applications have highlighted the “deviant” uses of verbs in specialised
corpora compared with general uses as well as the need for verb extraction. However, very few attention has
been given to verbs both in the terminology theory and automatic term extraction. In the light of a manual
corpus-based description, we investigate the status of verbs in an English-French (highly specialised) parallel
corpus and advocate a verb-oriented analysis in the framework of a corpus-based parser adapted to verb
extraction. Section 1 deals with the status of verbs in the terminology theory; section 2 introduces the framework
of the experiment and focuses on the characterisation of the parallel corpus in the domain of Computer Science.
Section 3 is dedicated to the corpus-based manual description. Finally, section 4 introduces a corpus-based
automatic analysis.
1.
Status of verbs
In the terminology theory, the status of verbs has been put aside for a long time. Only in recent years have
terminologists and lexicographers started to study the issue of defining verbs as terms for lexicographic
descriptions, term base creation, or ontology building. Therefore, studies of verbs as terms have been sparse until
now. However, the need for categorising verbs as terms has slowly emerged with the growing use of electronic
corpora in term extraction and phraseology description. Potential applications of term bases have raised the
hypothesis that terms were not only nouns, as the rule was in the Wüster (and domineering) approach. A "textual
terminology" approach, developed by (Bourigault and Slodzian 1999), and based on the use of electronic
corpora, opened the way for such questioning.
Descriptions of Language(s) for Specific Purpose(s) (LSPs) hardly consider the status of verbs. This is
particularly true of the Computer Science (CS) field - one of the most frequent LSP taught in France Numerous
textbooks provide students with a description of CS English. However, the description focuses on noun
terminology – multi-word nouns are for instance widely described - and on specific grammatical features used in
CS English, such as the passive or comparison. Very few attention is dedicated to the sentence, i.e. to verbs and
their distributional and transformational properties. Specialised dictionaries or bilingual glossaries mostly focus
on nouns. When verbs are mentioned, they are usually described as derived from nouns, with no other
information than part of speech. Bilingual glossaries are likely to give translations of verbs, but with very little
syntactic and semantic information on how to use them in a sentence.
The need for describing verbs as terms has been experienced in various applications. (Kübler and Foucou (to
appear)) show that French learners of CS English face major comprehension and production problems that are
very often related to specialised verbs. Building corpus-based teaching applications of CS English requires a full
description of the specialised verbs of the domain. (Kübler 2002) shows the need for verb description in
specialised translation, especially when machine translation is used by non-programming translators. This need
raises the issue of sorting out "verb terms"1 from general verbs in specialised texts. In the lexicographic area,
(L'Homme 1993, 1998) tried to forge definition criteria for verbs as terms in the French general CS domain. The
first criterion consists in considering that if the arguments of the verb are terms, there is a high probability that
the verb is a term too. This criterion is however based on the a priori characterisation of a term. (Pearson 1998)
suggests that the context is the only way of making the difference between a term and a word. (Frérot 2001) has
shown that arguments belonging to the general lexicon could indicate the presence of a verb term, by collocating
with the verb, revealing a particular behaviour which is different from its general use. The second criterion
defined by L'Homme considers verbs that are related to other lexical units that have already been identified as
terms. The issue of the a priori term determination remains unsolved. Furthermore, verbs and nouns terms that
are morphologically related are not always terminologically related. The analysis undertaken in this paper takes
those criteria into account, but goes further in considering the syntactic differences the verb terms exhibit
compared with their general use. The syntactic criterion has revealed to be most important as general language
intermingles with LSPs in specialised texts. Most verbs that can be found in the general language have specific
syntactic behaviours in LSPs and are subject to very specifically defined semantic restrictions (Frérot 2001).
1
We will use the "verb term" expression to refer to verbs that can be considered as terms.
429
Consequently, both verb analysis and description are crucial and terminology acquisition tools are therefore
expected to yield verb-related results. In designing Syntex, a corpus-based parser used to generate lexical
resources from specialised corpora, (Bourigault and Fabre 2000) have taken it into consideration and have
extended automatic extraction to verbs and verb phrases. Indeed, a verb-oriented analysis can improve
terminology extraction as it helps to better identify syntactic dependencies in the sentences of a corpus as well as
it enhances the distributional analysis (i.e. the grouping of words and phrases appearing in similar syntactic
contexts) used for the construction of semantic classes.
2.
Framework of the experiment
Corpus linguistics has now proven to yield reliable – and sometimes surprising – linguistic information.
However, the need for automating linguistic information extraction has increased, as globalization raises the
issue of multilingual document processing. Term extraction will thus become a key issue for the language
industry, helping to build multilingual databases or ontologies for the semantic Web applications.
Term extraction not only deals with well-written texts, revised and corrected by language professionals. It
will have to deal more and more with "naturally-occurring" texts, showing ill-formed language and various
genres. For this reason, we decided to choose a corpus that was very specialised, and provided a good example
of "naturally-occurring" texts: the Linux HOWTOs, which are the "user manual" of the Linux operating system.
Those texts have not been written by language professionals and are highly specialised. They have been written
in English and translated into many languages, among which, French. Aware of the needs for bilingual (but also
multilingual) terminologies, we worked on the English and French versions. The sample we used consisted of
200,000 words in either language. To check for uses in case of doubt, we also used the full HOWTO corpus,
which means ca. 500,000 words in each language; the Internet RFCs (ca. 8.5 million words) were also used to
double-check some uncertain uses. To make sure a verb structure was "deviant", i.e. specific to the subject area
compared with the "general" language, we also used French and English newspapers (one year of The Herald
Tribune and of Le Monde).
2.1. Characterisation of the parallel corpus in Computer Science
The HOWTOs are highly specialised texts written by CS experts, who are not always English native speakers.
The documents are aimed at CS experts2, and therefore do not address a wide audience. The communication
context can be that of an expert to expert communication and share the following parameters:
“It is assumed that author and reader share a common language and that when certain words or
phrases are used, each understands what is meant […]. Writer and reader, or speaker and hearer, are
assumed to have the same level of expertise […]. This expert to expert communication context is likely
to be the one with the highest density of terms” (Pearson 1998).
The HOWTOs show therefore a high percentage of terms and present features that are specific to very
specialised CS texts, such as command names, code lines, URLs,e-mail addresses, etc.. Below is an example of
the type of highly specialised sentences that can be found :
EN:
FR:
Java API with c-tree Plus' ISAM functionality gives Java functionality through native methods/RMI.
L'API Java associée aux fonctionnalités ISAM c-tree Plus permet des fonctionnalités Java au travers
de méthodes natives/RMI.
The English HOWTOs – the source texts – do not show the consistency and quality one can find in texts
written by technical writers. The syntax is sometimes shaky, as is the idiomaticity of some documents written by
non native speakers. No guideline, such as simplified English, has been used to avoid ambiguïties , such as e.g.
cable and ADSL connection, instead of cable connection and ADSL connection. The French translations are not
made by language professional either, but by experts in the field. The translations are not consistent with each
other, providing thus several different translations for the same use of one term. As the translations are made by
different people who are not professional translators, they can widely differ. As is often the case among French
computer scientists inside their work environment, English terms are used instead of the recommended French
translation (cf. expert FR: disquette de boot for EN: boot disk, instead of FR: disquette d'amorce). Another
common feature consists in using the English verb and adding a French verbal suffix, instead of using the French
equivalent, as the following examples show :
English verb term
to boot
to telnet
French verb term equivalent
démarrer/amorcer
se connecter par telnet
2
English verb with French suffix
booter
telneter
One must be familiar with Linux, or at least UNIX systems, to really understand the HOWTOs and make best
use of those.
430
This corpus is a good example of what must be really dealt with, i.e. texts that more closely reflect the
performance aspect of the language, than the competence of the ideal hearer-speaker. Although analysing the
HOWTOs do not really mean studying the performance of the "speakers" – the circumstancial characteristics of
the communication situation are lost in written texts3 – some typical performance features can be detected, such
as typos4. However, ill-formed words or sentences, and the use of various types of translations for the English
terms, cannot be explained only by performance mistakes, Hymes5 approach of communicative competence is
best adapted here to justify the non-standard linguistic structures and translations that are observed. Any
application, based on the chomskyan definition of competence will not be able to correctly deal with this kind of
texts.
2.2. From corpus-based manual description to automatic analysis
As mentioned above, term extraction has become a key issue in language industry, because of the increasing
need for corpus-based multilingual databases, or ontologies. Corpus-based manual description have allowed
linguists and language professionals to unveil actual language behaviour, and to dispose of real and statistical
language data. However, corpus-based manual description requires an investment in time and energy that
industry cannot afford. Hence the raising need for automating processes that have been tested by linguists.
Section 3 describes the methodology applied for corpus-based manual description, highlighting the value of the
resulting linguistic information, but also the weight of such a heavy and time-consuming task.
3.
Corpus-based manual description
3.1. Methodology
Using a concordancer allowing perl-like regular expressions on a corpus that was not POS tagged, we extracted
all entries that could be verbs. Heuristics were applied to extract verb candidates, such as words ending in –ing
or –ed, preceded by auxiliaries, modals and to. A concordance was then processed for each verb in English. As
the English and French HOWTOs are aligned in the concordancer interface, we studied the French equivalents
for each English verbal occurrence. As will be shown, one English verb term may have several French
translations that not always depend on different uses, but also vary depending on the translator. The
concordances allowed us to analyse the syntactic contexts of the verbs identifying the arguments in the different
syntactic positions, in order to build semantic classes.
The theoretical and methodological approach we used as a tool to analyse verb structures is based on the
lexique-grammaire6, which describes each verb via a basic sentence and divides them into classes, according to
their basic structures and their common transformational and distributional features. However, corpus
observations show that this approach does not take into account some particular features, such as syntactic
constraints, which restrict the cooccurrence of arguments. Moreover, a general syntactic description using basic
semantic features, such as human, abstract, place etc., is obviously unsatisfactory for the description of LSPs,
whatever the use of the description is. The classes of arguments that take the different syntactic positions must be
described extensively, hence the necessity of automatic term extraction. However, using regular expressions, we
tried to extract lists of potential arguments for verbs. Below is an example of extraction for the verb to run that
shows that there is still noise in the result:
run\w* \w+[^\.] \w+[^\.] \w*[^\.] (?:on|under|at) \w+ .{0,30}
run fdisk or cfdisk on it for you. Of the two, cfdisk is
runs to control everything on a machine, AND one is run per
run the following program on the client: #include <stdio.h> #
run the serial interface at a FIXED speed whilst allowing
run your NIS slaves on a Linux box? Or perhaps your
The manual description made use of comparison with general language (The Herald Tribune + Le Monde), in
order to check the degree of specialisation of some verbs, using thus the contextual criterion to decide on the
term status of verbs. Although newspapers are not completely representative of the language in general, they are
general enough to give good hints on the degree of specificity of a term.
3.2.
Some results
3
such as hesitations, cuts, repeated segments of a sentence that are typical of spoken situations.
such as (a) characters inversion, (b) missing character, or (c) replaced character, e.g. (a) *vebr, (b) *veb, or (c)
*vern for verb.
5
see Hymes D. (1972) On Communicative Competence. in J.B. Pridde and J. Holmes (eds) Sociolinguistics:
Selected Readings. Baltimore: Penguin
6
see Gross M. (1975) Méthodes en syntaxe. Paris: Hermann.
4
431
Let us take the description of the verb to run. Reference manuals give very few indications on its various uses.
Dictionaries of computing 7 do not mention it. The Merriam-Webster's gives only one use which is related to
computing : to run a problem through a computer. This use is also mentioned in the Collins-Cobuild, but along
with another one : You don't need a degree in mathematics to run (= operate) a computer. A quick check in the
HOWTOs and RFCs corpora yielded only four occurrences of run something through in the HOWTOs, and none
in the RFCs. Moreover, the arguments of to run do not match with the ones found in the dictionaries :
Dictionaries :
Corpus :
To run a problem through a computer
If you run your file through TeX program
Bilingual dictionaries gave us the following translations : exécuter, passer, fonctionner, être en marche, and
utiliser. Manually analysing the occurrences of to run in the corpus showed us that there are other translations in
use. The information that could not be found in the dictionaries included syntactic and semantic properties.
Translation inconsistencies are also noticeable in the corpus, as different French verbs are used for the same
English term, without any linguistically motivated reason.
To run shows a basic syntactic structure with three arguments, that does not exist in general English. Two
French equivalents are possible:
N0 runs N1 Prép N2
<=> N0 (lance + exécute) N1 Prép N2
N0 =: Nhum + applications that boot, such as LILO
N1 =: command name + programme
N2 =: machine, platform + programme + operating system
(1)
as the ability to run different programs in different virtual terminals
comme la possibilité de lancer des programmes différents dans différents terminaux virtuels
(2)
It just runs a command, which could be any Linux sound system
Il ne fait qu'exécuter une commande qui pourrait être n'importe quel programme de son sous
Linux
(3)
So you write 32-bit code that runs in 16-bit mode on a 32 bit CPU.
vous écrivez donc du code 32 bits, qui s'exécute en mode 16 bits sur un processeur 32 bits.
Arguments can change depending on the preposition:
N0 runs on N1, with : N0 =: applications + operating system (Linux, Win95, X-Window), N1 = : machine,
platform (PC, 21066,) + operating system
(4)
VirtuFlex runs on standard Unix Workstations
VirtuFlex tourne sur des stations Unix standard
N0 runs under N1, with : N0 =: applications, N1 = : operating system
(5)
ANS FORTH system that successfully runs under Win32s, Win95, Win/NT
système ANS FORTH 32 bit libre qui fonctionne sous Win32s, Win95, Win/NT
N0 runs (at +with) N1, with N0 = : Nhum, N1 = : applications, Prep = :(at + with), N2 = : N-hum
(6)
Generally, the PCI runs at 33MHz
En général, le PCI tourne à 33MHz
A causative construction is possible in French, with the introduction of the operator faire :
(7)
You can run Linux on any Alpha-based machine
Vous pouvez faire tourner Linux sur n'importe quelle machine Alpha
Comparing those examples with general English uses can help isolate technical contexts. The structures
described above do not exist in general English; on the other hand, there are structures in general English that
cannot be used in CS English, as in become a presidential concern about running for re-election in 1996 or
stamps, old coins and odd documents, run around the square. The same does not apply to French tourner as in
quatre poules blanches tournant en rond sur une place de village, since the structure is similar to the specialised
use. There are other examples in general French that are not in use in CS French, such as Quant au cachet de
Barbra Streisand, il tourne autour de 20 millions de dollars. The drawback of this kind of description is that it
does not explicitely shows the different structure combinations that are possible, as will be shown in section 4.
Another significant example is the verb boot. To boot is quite frequent in the corpus (around 700
occurrences). However, general and even specialised dictionaries give little information on this verb. In the online Merriam-Webster's8 only general uses of to boot can be found: to avail, to profit. The Collins-Cobuild offers
7
8
FOLDOC, A Glossary of Computing Terms, Dictionary of Computing For Learners of English
http://www.m-w.com
432
no verb entry for to boot. Wordnet9 provides some information on the specialised use of the verb (n°2 below),
but with very little syntactic semantic information:
1. Boot : kick ; give a boot to
2. boot : cause to load (an operating system) and start the initial processes
The basic structure of the verb, as analysed in the corpus, has three syntactic positions that can be filled in by
specific arguments. The subject is the agent of the action:
N0 boots N1 Prep N2, with the following argument classes :
• N0 =: Nhum or applications such as LILO which work as a metaphor, as they can be attributed the agent role.
• N1 = : operating system, system, disk, bootdisk, hard disk, floppy disk, kernel => all bootable objects
• N2 = : CD, CD-ROM, D :, C :, A :, file, emergency disk => booting objects
Three prepositions are possible with that structure: off, with, from. An idiosyncratic use of the phrasal verb to
boot off has been detected (4) :
(1)
(2)
(3)
(4)
To boot one of your old kernels off the hard drive…
Pour lancer l'un de vos vieux noyaux à partir du disque dur…
A good idea might be to boot the notebook with a kernel
Une bonne idée serait de démarrer le portable avec un noyau
In order to have LILO boot Linux from OS/2 Boot Manager,
Afin que LILO lance Linux à partir du gestionnaire de démarrage d'OS/2,
You can boot off of a floppy disk
Vous pouvez démarrer à partir d'une disquette
PP deletion, allowing a transitive sub-structure with arguments restrictions can be found. As to boot is an
ergative verb, an intransitive structure in which the subject argument is can be analysed as the patient affected by
the action is allowed. In this case, the French equivalent is a pronominal structure, which is very often used to
translate English passives:
(5)
When Linux boots, it is usually configured not to produce…
Quand Linux se lance, il n'est habituellement pas configuré pour produire…
An intransitive structure with to, and into has been found in the corpus:
(6)
Your BIOS may not allow you to boot directly to a SCSI drive.
Votre BIOS ne vous permettra peut-être pas de démarrer directement à partir d'un disque SCSI
The syntactic and semantic properties described above show the difference between the neologism to boot and
the general verb, that has no etymological relationship with the specialised one. The general verb behaves very
differently. Here are two examples extracted from one year of the Herald Tribune that speak for themselves:
In early 1988 the Saudis booted out Hume A. Horan
eating habits under control by booting the French chef and his staff.
French examples extracted from the "general" corpus of Le Monde show immediate differences in the use of
lancer (as a translation of to boot):
et l hymne fraternel que lance à ce dernier Jérôme Garcin.
MCI se lance dans la bataille des " autoroutes de
Martine Aubry lance le débat sur le partage du temps de travail
The last verb we will exemplify for the corpus-based manual description here is to dump. The Robert & Collins
Super Senior English-French dictionary gives the following examples and French translations (which are not in
the corpus) dump (Comput) data, file, etc vider – to dump to the printer transférer sur l'imprimante.
To dump shows transitive locative constructions10, such as:
N0 dumps N1 on N2, with the following argument classes :
(1)
a fast computer […] can dump 32k of data on you
qu'un ordinateur rapide […] pourra vous inonder de 32ko de données
N0 dumps N1 (to + onto) N2, with the following argument classes :
• N0 =: Nhum.
9
http://www.cogsci.princeton.edu/~wn/
Guillet A. et Leclère Ch., 1992. La structure des phrases simples en français: tome II: constructions
transitives locatives . Genève: Droz.
10
433
•
•
(2)
(3)
N1 = : data, content, memory
N2 = : disk, file
dumps its memory image to disk in executable format
écrit l'image de sa memoire sur le disque sous format binaire
you just dump the contents of one disk onto the other
en copiant directement le contenu d'un disque sur un autre.
The different French translations account for two different uses depending on the preposition (on or (in)to). A
query on the Herald Tribune corpus shows that the specialised uses are quite different from the general ones, i.e.
no occurrence of the preposition to :
for defying his demands to dump her boyfriend and cut her long hair
overseas investors will dump Japanese stocks
The storm, which dumped 23 centimeters (9 inches) of snow Saturday in Tokyo,
The same applies to French. General uses of écrire and copier show significant differences as far as structures
and argument classes are concerned:
Ecrire sur la mort d un ami est périlleux.
Réalisé avec de gros moyens, ce film copie Alien sans vergogne.
alors elles nous copient sur les spécialités!"
In this section, the corpus-based manual description highlighted idiosyncratic uses of verbs in our corpus.
Even though we used aligned concordances to carry out the analysis, the procedure – as it is not fully automated
– was time-consuming and quite hard-working. However, this methodology proved useful to highlight the uses
that are specific to CS and do not belong to general language. Therefore, we investigated the possibility of using
a corpus-based automatic tool - both for French and English - that would yield verb-related results, and above
all, relies on the corpus as much as possible. Section 4 is thus dedicated the a corpus-based automatic analysis.
We first point to a few arguments in favour of verb analysis and then give a brief overview of the automatic tool
we used. Finally, we show - through Prepositional Phrase (PP) attachment - how endogenous techniques are
well-adapted to our verb study.
4.
Corpus-based automatic analysis
4.1. A few arguments in favour of verb analysis
As mentioned in Section 1., both verb analysis and description are crucial and terminology acquisition tools are
therefore expected to yield verb-related results, though most tools do not go as far as verb output. Let us insist on
three (at least) reasons accounting for a verb-oriented analysis.
Reason 1. From a terminological point of view, verbs may be terms just as nouns do, following (Bourigault
and Jacquemin 2000) who postulate a verb terminology. Indeed, verbs play a major role in specialised texts and
are likely to be terms on the basis that they can be: (i) morphologically related to a noun or noun phrase being
itself a term such as the French and English examples11: injecter un virus / injection d'un virus, to access the
system / an access to the system; (ii) specialised verbs (usually simple verbs), i.e. verbs whose use is restricted to
a given domain and refer to a specific concept, such as the French verbs transduire or transfecter in the domain
of gene therapy, or the English CS verbs to telnet, to bufferize; (iii) verbs exhibiting a “deviant” use in the
terminological system in comparison with their expected use in the lexical system. In this context, “deviance12”
refers to the unpredictable verb argument structure both in terms of syntactic and semantic behaviour (Frérot
2001). Let us illustrate that point with the following corpus-based examples : construire des souris, recruter des
cellules, the daemon listens to all the messages.
Reason 2. A verb-oriented analysis improves automatic term extraction as it helps to better identify the
constituents (frontiers) of sentences in a corpus, therefore increasing noun extraction accuracy. Let us look at the
following examples:
(a) to boot [a Linux kernel] [on a CD ROM]
11
The corpus-based examples in section 4 are taken from the French-English HOWTO corpus and French
corpora in the domain of geomorphology and gene therapy.
12
« […] we postulated deviance as the linguistic characteristic of terms in relation to words. The deviance was
described as being of several kinds ; 1. Unsually high frequency of compound verbs. 2. Coinage of new words. 3.
Unusual syntactic behaviour : new or forbidden constructions. 4. Unusual semantic behaviour : appareance of
new meanings which show themselves by unsual combinations » (Condamines 1995).
434
Æ *13 Linux kernel on a CD ROM
Æ to boot on a CD ROM / Linux kernel
(b) to give [the compiler] [hints] about how to optimize
Æ *compiler hints
Æ to give hints / compiler
(c) enrober [de calcite] [des matériaux]
Æ * calcite des matériaux
Æ enrober de calcite / matériaux
In the above examples, a noun focus alone will not allow a proper analysis of the sentence, as in (a) it leads to
identify as a potential term Linux kernel on a CD ROM, though the preposition on depends on the verb boot ; the
same applies to (c) : des matériaux is the direct object of enrober, the determiner des depends on the verb
enrober (and not calcite). In this context, adopting a verb approach obviously reduces the generation of invalid
terms (*Linux kernel on a CD ROM , *calcite des matériaux), its correlate being the necessity to deal with verb
structures.
Reason 3. Syntactic verb contexts are very productive for the distributional analysis as they enhance the
grouping of words and phrases appearing in similar syntactic contexts, which is used in our tool for the
construction of semantic classes.
to produce
Æ {basalt, crust, flow, lava, magma}
to generate
In the above example, the nouns basalt, crust, flow, lava, magma are said to form a cohesive semantic class, on
the basis that they share similar contexts with the verbs produce and generate.
In designing Syntex, a corpus-based parser used to generate lexical resources from specialised corpora,
(Bourigault and Fabre 2000) have taken those three parameters into consideration and have extended automatic
extraction to verbs and verb phrases.
4.2. Syntex : a corpus-based parser adapted to verb extraction
Syntex is a corpus-based parser14 used to generate specialised lexical resources, such as lexicons for translation,
ontologies or thesauri, and has been used in various « real world » applications (among the most recent are
(Bourigault and Lame 2002, Le Moigno et al. 2002, Chodkiewicz et al. 2002)). Syntex first15 identifies lexicosyntactic dependencies in the sentences of a given corpus (for instance, subjects, direct or indirect objects of
verbs) and builds a network of words and phrases in which each phrase is linked to its syntactic heads and
expansions16. The network is then used as a material for the construction of semantic classes on a distributional
basis i.e. the grouping of words and phrases appearing in similar syntactic contexts (for an accurate description
of the distributional analysis module, see (Bourigault 2002)). We will focus here on the extraction of lexicosyntactic dependencies - with an emphasis on PP attachment - as they are the starting point in the whole process
and will give an overview of the general principles underlying the analysis.
Syntex's major specificity is to rely on endogenous techniques (Bourigault 1994) which allow the parser to
acquire, for every new corpus analysed, the subcategorization information necessary to resolve syntactic
attachment ambiguity. This strategy is based on in-depth studies of various domain corpora, highlighting
idiosyncratic uses of lexico-syntactic structures compared with their general use and from one domain to another
(Fabre and Bourigault 2001, Basili et al. 1997, Basili et al. 1999). In this context, using general linguistic
knowledge tends to prove quite inefficient and irrelevant. Consequently, Syntex does not use any a priori
linguistic resources. Let us illustrate the endogenous strategy on PP attachment, the very first procedure in
delimitating phrases.
- English Example :
Verb
Noun
Preposition
13
* indicates a wrong analysis (invalid term).
There is a French version of Syntex as well as an English version.
15
Before Syntex is used, the corpus is morphosyntactically tagged (each word in the corpus is assigned a lemma
and grammatical tag).
16
Example : in the French noun phrase plan de faille, plan is the head and faille the extension.
14
435
Ambiguous case :
to run programs in virtual terminals
Æ potential governors for the preposition in :
? in
program
run
In order to solve this ambiguous case, in other words find the preposition's governor, our tool relies on
unambiguous cases, i.e. cases containing only one governor, such as :
Unambiguous cases :
Dosemu has to be run in another terminal
compile and run in double precision mode
why not run in a Linux box your NIS slaves
Syntex relies on those unambiguous cases to compute productivity measures used to perform the right
attachment :
(run, in {terminal, mode, box}) : corpus-based occurrences
(run, in) : productivity = 3
Æ to run different programs in virtual terminals
We now illustrate this procedure with a French example.
- French example :
Æ introduire du matériel génétique dans les cellules
ce qui permet de les introduire dans une cible
a introduit dans une cellule eucaryote de l'ADN
du matériel génétique a été introduit dans l'organisme
(introduire, dans {cible, cellule, organisme}) : corpus-based occurrences
(introduire, dans) : productivity = 3
Æ introduire du matériel génétique dans les cellules
4.3.
First experiment on the HOWTO corpus
4.3.1 A few remarks on pre-processing
As mentioned earlier Section 2.1., the HOWTO corpus is characterised by a vast number of features such as
URLs, e-mail addresses, code names, command lines or scripts - among others - that make the task of preprocessing difficult (by pre-processing, we mean sentence and word segmentation as well as morphosyntactic
tagging). For instance, the high number of enumerations, in the form of listing, adds to the difficulty of sentence
segmentation which becomes even more complex with the “naturally-occurring” dimension – referring to the
lack of homogeneity in punctuation, typography, due in part to different people writing the user guide (be it in
English or for the French translations).
As far as the morphosyntactic tagging is concerned, the tagset used covers part of the corpus specificities the NomMail and NomUrl tags are well-suited to analyse phenomena such as URLs or e-mail addresses
(examples : NomMail|[email protected] / NomUrl|http://www.eleves.ens.fr) - though it does not cover
all of them. Let us mention also a few more general17 tagging errors, regarding for instance, ambiguous ing
forms in English. Taken as a whole, it should be pointed out that the pre-processing quality - from sentence
segmentation to morphosyntactic tagging - obviously impacts on the lexico-syntactic analysis.
4.3.2 Verb parsing output
We will now more closely look at the verb analysis - through PP attachment - performed by the parser and
emphasise on an English verb sample {boot, compile, configure, dump, mount, run} which we believe to best
represent the endogenous procedures as well as the type of verbs found in specialised corpora. In Section 3., we
showed that the manual corpus-based description highlighted idiosyncratic uses of those verbs compared with
their general uses, thus implying the need to use automatic tools that adapt to corpora. Therefore, what we intend
to show here is that endogenous procedures are particularly well-adapted to corpora as they “respect” their
specificities and use no other information but that of the corpus.
As mentioned earlier., our tool relies on endogenous techniques and more precisely on the corpus
productivity, implying lexico-syntactic redundancy. Let's analyse the verb run in the corpus, which gets
constructed with a wide variety of prepositions as shown below:
17
General, as opposed to specific to the HOWTO corpus.
436
ability to run different programs in different virtual terminals at the same time
other subsystems (DRAM, for example) will run asynchronously at lower clock rates
you can run any other program from within emacs
can run the PCI at any frequency
the above tests were run with some of the special Cyrix
Run dosemu with partition access
your application which would probably run under the IBCS2 emulator
to run dosemu inside a color xterm
Æ (a) and runs on all platforms
Æ (b) VirtuFlex runs on standard Unix workstations with 8 MB of RAM minimum
Æ (c) why not run you NIS slaves on a Linux box
Æ (d) to run popular Windows applications on Linux based system software
Among all the verb occurrence frequency, run is one of the highest in the corpus, with approximately 400
occurrences. The more occurrences in the corpus, the more chances for our tool to find unambiguous cases (a)
(b), used to resolve ambiguous cases (c) (d). The same remarks apply to the following verbs {compile, configure,
dump, mount} which get constructed with various prepositions and whose number of occurrences (respectively,
224, 190, 50, 115) is high enough to allow the parser to resolve PP attachment ambiguity. The table below shows
the verb-preposition associations and highlights some idiosyncratic uses and missing dictionary-based
descriptions, such as dump back to, compile in support, compile in support for, cross compile from-to, which the
manual verb description had indeed revealed.
Verb
Preposition
compile
configure
dump
mount
Occurrences
against, as, from-to, into, to compile things against this library, to compile as ELF, to cross compile from
with, in, in support, in Linux to Dos, the steps to compile into the kernel, the driver was compiled with
debugging enabled, to compile in your own personal values, to compile in support
support for, out of
to use the program selection, you need to compile in support for your CDROM
drive, it compiles right out of the box on Linux
as, for, in, in support for, configuring your machine as an NCP server, the card is configured for shared
on, to, under, with,
memory operation, allow the card to be configured in software, do not configure in
support for the 82C710, devices to be configured on a Linux machine, the card is
not configured to one of the addresses, configure CDROM drive under Linux, you
have configured dosemu with a command like $
on, into, onto, to, back to you can dump 32k of data on you without stopping, it will dump everything into a
ftape-2., when you just dump the contents of one disk onto the other, dump the
image to the disk, dumping its memory image back to disk
as, for, from, into, on, name of the directory to mount as root, to mount a CDROM for read/write, a
through, under, with,
CDROM is mounted from Linux, to mount Novell volumes into your Linux
filesystems, mounting the CDROM on bootup, mount the empty files through the
loopback devices, you dos partition is assumed to be mounted under Linux, the root
filesystem is mounted with write access
The automatic analysis of the French translations for {boot, compile, configure, dump, mount, run} showed
very frequent verb-preposition associations - be they long distance or adjacent - , such as for boot :
lancer/démarrer {à partir de, dans, sur} or run : faire tourner/exécuter {dans, sous, depuis}. Our endogeous
procedures proved efficient18 to resolve the PP attachment ambiguities. Syntactico-lexical redundancy in the
corpus was sufficient for the parser to correctly attach the preposition to the verb. Finally, it should be pointed
out that so far, no distinction is made in the parser between arguments and adjuncts of verbs. Whatever the status
of PPs - not as clear-cut as often claimed -, they are attached to the verb. This choice is based on the assumption
that non argumental relations between verbs and PPs highly contribute to word semantic neighbouring, hence to
word semantic class. More generally speaking, corpus-based tools may highlight and bring new insight to
linguistic phenomena - as is the case here for the argument/adjunct distinction – that an intuition-based manual
description, strongly influenced by categorical models of grammar would not show.
5.
Conclusion and future work
This first experiment has shown the necessity of using corpus-based manual description as an incentive for
automatic term extraction. Manual descriptions in LSPs highlight the "deviant" uses that are not found in the
general language. However, manual corpus work being time-consuming and since the need for exhaustive
18
We exclude wrong analyses due to tagging errors.
437
linguistics description has increased, automating linguistic information extraction will enable linguists and
language professionals to go further in research and applications. The approach implemented in Syntex indeed
takes into account the LSPs specificities highlighted by corpus-based manual descriptions. As the use of machine
translation has developed in the recent years, bilingual term extraction will be more and more needed to build
MT specialised dictionaries, in order to improve translation results. Verb term extraction can also prove very
useful in computer-assisted language learning (CALL), leading to automated exercise generation and helping in
the correction process. Future work will deal with the complete bilingual extraction of verb terms from the
HOWTO corpus, in order to test the creation and validity of MT dictionaries.
References
Basili R, Pazienza M-T, Vindigni M 1999 Adaptative Parsing and Lexical Learning. Actes de VEXTAL'99,
Venise.
Basili R, Pazienza M-T, Vindigni M 1997 Corpus-driven Unsupervised Learning of Verb Subcategorization
Frames, Actes du 5ème congrès AI*IA 97, M. Lenzerini (ed), Lecture Notes in Artificial Intelligence, 1321, pp.
159-170.
Bourigault D 1994 Lexter, un logiciel d'extraction de terminologie. Application à l'acquisition de connaissances
à partir de textes, Thèse de l'Ecole des Hautes Etudes en Sciences Sociales, Paris.
Bourigault D 2002 Upery : un outil d'analyse distributionnelle étendue pour la construction d'ontologies à partir
de corpus. Actes de la conférence TALN, Nancy, 75-84.
Bourigault D, Jacquemin C 2000 Construction de ressources terminologiques. In : J-M. Pierrel, Ingenierie des
langues, Paris : Hermès Sciences Publications, Chap. 9 : p. 215-233.
Bourigault D, Fabre C 2000 Approche linguistique pour l'analyse syntaxique de corpus. Cahiers de grammaire,
Vol.25, pp.131-151.
Bourigault D, Slodzian M 1999 Pour une terminologie textuelle. Terminologies nouvelles 19: 29-32.
Chodkiewicz C, Bourigault D, Humbley J 2002 Making a workable glossary out of a specialised corpus: Term
extraction and expert knowledge, in Altenberg B & Granger S. (eds), Lexis in contrast, corpus-based
approaches, John Benjamins Publishing Company, Amsterdam/Philadephia,, pp. 249-267.
Condamines A 1995 Terminology : new trends, new perspectives. Terminology 2:2, 219-238.
Fabre C, Bourigault D 2001 Linguistic clues for corpus-based acquisition of lexical dependencies. Actes de
Corpus Linguistics Conference, Lancaster, pp. 176-184.
Frérot C 2001 Caractérisation du verbe en terminologie. Application au domaine de la thérapie génique en
cancérologie. Mémoire de DEA, Université Paris 7.
Kübler N 2002 Teaching Commercial MT to translators: Bridging the Gap between human and machine. In H.
Somers (ed.) Proceedings of the EAMT workshop on MT, Manchester, UMIST. pp. 155-162.
Kübler N, Foucou P-Y. (to appear) Teaching English Verbs With Bilingual Corpora : Examples in the Computer
Science Area. in S. Granger &S. Petch-Tyson (ed) : Contrastive Linguistics and Translation Studies, Rodopi,
Amsterdam.
L'Homme M-C 1993 Le verbe en terminologie : du concept au contexte. L'actualité terminologique 26(2) 17-19.
L'Homme M-C 1998 Définition du statut du verbe en langue de spécialité et sa description lexicographique.
Cahiers de lexicologie 73(2) 125-148.
Le Moigno S, Charlet J, Bourigault D, Jaulent M-C 2002 Terminology extraction from text to build an ontology
in surgical intensive care, in Proceedings of the AMIA 2002 annual symposium (American Medical Informatics
Association), San Antonio, USA,
Pearson J 1998 Terms in Context, Amsterdam: John Benjamins Publishing Company.
438
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