Pre-print

Pre-print
Lang Resources & Evaluation
DOI 10.1007/s10579-014-9271-6
ORIGINAL PAPER
A qualitative comparison method for rhetorical
structures: identifying different discourse structures in
multilingual corpora
Mikel Iruskieta • Iria da Cunha • Maite Taboada
Received: 26 June 2013 / Accepted: 8 May 2014
Ó Springer Science+Business Media Dordrecht 2014
Abstract Explaining why the same passage may have different rhetorical structures when conveyed in different languages remains an open question. Starting from
a trilingual translation corpus, this paper aims to provide a new qualitative method
for the comparison of rhetorical structures in different languages and to specify why
translated texts may differ in their rhetorical structures. To achieve these aims we
have carried out a contrastive analysis, comparing a corpus of parallel English,
Spanish and Basque texts, using Rhetorical Structure Theory. We propose a method
to describe the main linguistic differences among the rhetorical structures of the
three languages in the two annotation stages (segmentation and rhetorical analysis).
We show a new type of comparison that has important advantages with regard to the
quantitative method usually employed: it provides an accurate measurement of
inter-annotator agreement, and it pinpoints sources of disagreement among annotators. With the use of this new method, we show how translation strategies affect
discourse structure.
Keywords Annotation evaluation Discourse analysis Rhetorical Structure Theory Translation strategies
M. Iruskieta (&)
Department of Didactics of Language and Literature, University of the Basque Country,
Sarriena auzoa z/g, 48940 Leioa, Spain
e-mail: [email protected]
I. da Cunha
University Institute for Applied Linguistics, Universitat Pompeu Fabra, C/ Roc Boronat 138,
08018 Barcelona, Spain
e-mail: [email protected]
M. Taboada
Department of Linguistics, Simon Fraser University, 8888 University Dr, Burnaby, BC V5A 1S6,
Canada
e-mail: [email protected]
123
M. Iruskieta et al.
1 Introduction
Translation or parallel corpora on the one hand and comparable corpora on the other
are useful in many tasks, in applied linguistics and in natural language processing.
Compiling such corpora can provide insight into translation strategies, can help
validate or disprove intuitions about differences across languages, and can be useful
in computational applications such as machine translation or terminology
extraction.
Translation corpora have been useful in testing hypotheses about language
contrasts. Granger (2003), for instance, using translation corpora, put into question
the over-generalization that ‘‘French favors explicit linking while English tends to
leave links implicit’’. Translation corpora also help identify strategies used in the
translation process, such as the strategy that Xiao (2010) found in translated Chinese
texts, where there was an increased use of discourse markers, presumably to more
clearly identify the rhetorical structure of the text (although introducing discourse
markers may lead to subtle changes in rhetorical structure as well, in cases when the
translator interprets a different relation than that intended by the original author).
Most contrastive corpus-based studies emphasize surface-level aspects of
language, such as differences in terminology in general (Gomez and Simoes
2009; Morin et al. 2007; Fung 1995; Wu and Xia 1994) and specific lexical items in
particular (Fetzer and Johansson 2010; Flowerdew 2010); differences in aspects of
modality (Kanté 2010; Usoniene and Soliene 2010); or the use of discourse markers
(Mortier and Degand 2009). There exists, however, a sizeable body of work on
differences in the rhetorical structure of texts across languages, in particular within
the framework of Rhetorical Structure Theory (RST), a theory of text structure
proposed by Mann and Thompson (1988). The first contrastive RST study
comparing one European language and one Asian language was carried out by Cui
(1986), who compared English and Chinese expository rhetorical structures. Kong
(1998) and Ramsay (2000, 2001) studied the same pair of languages, in both cases
examining specific genres (business request letters and news texts). Other pairs of
languages studied within RST include Arabic and English (Mohamed and Omer
1999), Japanese and English (Marcu et al. 2000), or a range of European languages,
such as Dutch-English (Abelen et al. 1993), Finnish–English (Sarjala 1994), FrenchEnglish (Delin et al. 1996; Salkie and Oates 1999), Spanish–English (Taboada
2004a, b), and Spanish–Basque (da Cunha and Iruskieta 2010).
Contrastive studies comparing the rhetorical structures of more than two
languages are not very common, although we can mention the study in Portuguese–
French–English by Scott et al. Scott et al. (1998). They show a methodology to
carry out RST contrastive analysis of instructional texts in different languages, and
they present the results of an empirical cross-lingual experiment based on this
methodology. More information about contrastive RST studies or studies about
other languages can be found in Taboada and Mann (2006a, b).
One observation in RST-based work is that the same passage, when conveyed in
two different languages, may have different underlying rhetorical structures
(Bateman and Rondhuis 1997; Delin et al. 1994). An explanation for such
differences is that translation strategies reorganize the structure of the discourse,
123
A qualitative comparison method for rhetorical structures
with the resulting underlying structures being different. Translation literature deals
with many aspects of this phenomenon, one being differences in explicitness, which
in some cases result in different underlying structures (House 2004).
This proposal (that translation strategies lead to different structures) is often
presented on the basis of individual examples, with no unifying principle for the
representation of underlying structure. In this paper, we present a new method for
the evaluation of discourse structures across multiple languages to analyze which
translation strategies affect rhetorical structure.
The first aim of this paper is to provide a new qualitative method to compare
rhetorical structures in different languages and/or by different annotators. Existing
work comparing different annotations uses a quantitative methodology (Marcu
2000a). The main comparison methodology consists of quantifying the agreement
between the rhetorical analyzes done by annotators, in terms of Elementary
Discourse Units (EDUs), spans (sets of related EDUs), nuclearity (nucleus or
satellite role of a span) and rhetorical relations (set of hypotactic and paratactic
relations). To compare rhetorical analyzes, typical precision and recall measures are
used. Work by da Cunha and Iruskieta (2010) and van der Vliet (2010) presents
some criticisms of Marcu’s methods, arguing that this quantitative method
amalgamates agreement coming from different sources, because decisions at one
level in the tree structure affect decisions and factors at other levels, with the result
that the factors are not independent. Disagreement on segmentation or attachment
point at lower levels in the tree significantly affects agreement on the upper
rhetorical relations in a tree, and should be accounted separately. Mitocariu et al.
(2013) have proposed an evaluation method (for RST and Veins Theory Cristea
et al. 1998) which checks the inner nodes1 (attachment point), nuclearity of the
relation (nuclearity) and the vein expressions or constitution of the units
(‘‘constituent’’ Marcu 2000a) but excludes the names of relations as a comparison
criterion. In our evaluation method we consider Mitocariu et al.’s factors
(attachment point, constituent and nuclearity) and the rhetorical relations. We
believe that the qualitative method that we present here addresses the deficiencies in
previous proposals and provides a qualitative description of dispersion annotation,
while at the same time allows the quantitative evaluation.
The second aim of this paper is to test this method. In order to detect differences
among rhetorical structures and study the origin of such differences, we analyze a
corpus of parallel texts in three different languages: English, a Germanic language;
Spanish, a Romance language; and Basque, a non-Indo-European language. We
investigate whether differences are motivated by different translation strategies or
by the choice of one relation over another in a group of similar relations, as Stede
(2008b) proposes. Our corpus, albeit small, is comparable to the only other
trilingual comparative corpus (Scott et al. 1998), and it is rich enough to allow the
development and evaluation of a qualitative comparison method for rhetorical
relations.
Our study is useful from a theoretical point of view, because it will help us
understand how the rhetorical structures of texts in different languages are
1
Soricut and Marcu (2003, pg. 152) use the term ‘‘attachment point’’ or ‘‘dominance set’’.
123
M. Iruskieta et al.
constructed. Moreover, the study provides rhetorical analyzes of a less-commonly
studied language,2 Basque, the only pre-Indo-European language of Western Europe
(Trask 1997) and one of the four official languages of Spain (together with Catalan,
Galician and Spanish), spoken in the Basque country. From an applied point of
view, this work supports the development of computational linguistics systems
(such as summarization, information extraction and retrieval systems), where
accurate annotation is of paramount importance. In addition, our methodology can
be useful in research on automatic compilation of specialized corpora, and can help
professional translators and machine translation researchers.
The paper is organized as follows: Section 2 presents the methodology and
theoretical background of our study. Section 3 describes our methodological
proposal and provides the results of the discourse analysis of our corpus. Section 4
provides conclusions and proposals for future work.
2 Methodology
Our work consisted of three stages. First, we decided on the theoretical framework
of our study, RST. Second, we built the corpus. Finally, we carried out the analysis,
including a comparison of the three different RST structures for each text, using
both a quantitative methodology and our proposed new qualitative methodology.
2.1 Theoretical framework
In this study, we use RST, since it is a language-independent theory. RST is a
descriptive theory for textual organization that characterizes text structure using
relations among the discourse or rhetorical elements that a text contains. These
elements are called spans, and they can be nucleus (if the element is more essential
to the speaker’s purpose) or satellite (if it provides some rhetorical information
about the nucleus). The relations can be: (a) nuclear relations (e.g., ANTITHESIS,
CAUSE, CIRCUMSTANCE, CONDITION, ELABORATION, EVIDENCE, JUSTIFICATION, MOTIVATION,
PURPOSE), that is, hypotactic relations between nuclei and satellites, and (b) multinuclear relations (e.g., CONTRAST, JOINT, LIST, SEQUENCE), that is, paratactic relations
among nuclei, where more than one unit is central with regard to the speaker’s
purposes. For a more detailed explanation of RST, see Mann and Thompson (1988)
and the RST web site by Mann and Taboada (2010).
RST relations are typically represented as trees. Figure 1 shows a fragment of an
RST tree,3 with one multinuclear relation (CONJUNCTION) and two multinuclear
2
Although great efforts have been made to stimulate Machine Translation studies for different language
pairs, non-official languages that are typologically different and could be interesting are not considered.
For example Koehn (2005) presents a 30 million word corpus translated to the 11 official of the European
Union: Danish, German, Greek, English, Spanish, Finnish, French, Italian, Dutch, Portuguese, and
Swedish to study different language pairs translations, but less common languages spoken in the EU are
not included.
3
The source of the text (TERM#_original language) is shown in square brackets at the end of the figures,
tables or examples.
123
A qualitative comparison method for rhetorical structures
Fig. 1 Example of an RST tree, TERM30_ENG
relations (RESULT and ELABORATION). The annotator recognized that spans 16 and 17
are conjoined, forming another span where each item has a comparable role
(moreover, each span has a verb are and appears, and they are linked by the
connector and). The annotator also found a RESULT relation, since she understood
that span 18 could be the cause for the situation explained into the span 19 (again,
each unit has a finite verb: is associated and [is] given, and they are linked by the
double connector and thereby). It is important to observe that rhetorical relations are
applied recursively, i.e., spans that stand in a relation: 18 and 19 in Fig. 1 form a
new span (18–19) that can enter into new relations, such as the ELABORATION relation.
In this case, the annotator labelled this relation as such because the span made up of
units 18–19 (satellite) provides additional information about the previous span (16–
17), which constitutes the nucleus of the relation. Following Marcu’s (2000b) strong
compositionality criteria, the most important units for the 16–19 span are 16 and 17.
For the span 18–19 the most important unit is 18.
In the literature on RST, there is agreement that the most important unit of the
tree is the ‘‘central unit(s)’’ (Stede 2008b) and the most important unit of a span is
the ‘‘central subconstituent’’ (Egg and Redeker 2010). So following this framework
we will use the term ‘‘Central Unit(s)’’ (CU) of the text for the most important unit
of an rhetorical structure tree (RS-tree) and ‘‘Central Subconstituent(s)’’ (CS) of a
relation for the most important unit of the modifier span that is the most important
unit of the satellite span. When there is a simple constituent (that is no more than
one EDU), we formalized this simple constituent as the CS, and when there is a
multinuclear relation, we describe it with all of its constituents.
Table 1 provides a representation of this example.
There are several classifications of RST relations: the classic one by Mann and
Thompson of 24 relations (Mann and Thompson 1988), the extended one by Mann
and Thompson of 30 relations, available on the RST site (Mann and Taboada 2010),
and Marcu’s classification of 78 relations (Carlson et al. 2003), among others. We
have chosen the extended classification for the annotation of our trilingual corpus.
Space constraints preclude an extensive discussion of its merits over other
approaches (see Taboada and Mann 2006a, for a discussion).
123
M. Iruskieta et al.
Table 1 Formalization of Fig. 1, TERM30_ENG
Relation
Left span
Right span
CS
Nuclearity
Result
18
19
19
NS
Conjunction
16
17
16–17
NN
Elaboration
16–17
18–19
18
NS
2.2 Corpus
As Granger (2003) proposes, a multilingual translation corpus is:
½. . . the most obvious meeting point between CL (Contrastive Linguistics) and
TS (Translation Strategies). Researchers in both fields use the same resource
but to different ends: uncovering differences and similarities between two (or
more) languages for CL and capturing the distinctive features of the
translation process and product for TS.
(Granger 2003, pg. 22)
In translation studies, where the intention is to search for similarities and
differences in large corpora, it is difficult to find a balanced corpus in size and
similar composition of genres (Baker 2004). Our problem was to find a balanced
multidirectional corpus of such size that allowed for a manual comparison of all the
rhetorical structures by language pair. One of our aims, as we said, is to propose a
methodology to describe when a different RST relation can be attributed to
annotator interpretation or to different language forms.
As far as we know, no multilingual corpus with English, Spanish and Basque
texts exists. Our corpus was then compiled specifically for this work.4 It is a
multidirectional translation corpus which contains abstracts of research papers
published in the proceedings of the International Conference about Terminology
that took place in Donostia and Gasteiz in 1997 (UZEI and HAEE-IVAP 1997). In
this conference, authors were allowed to send full papers in English, French,
Spanish or Basque, but they had to provide titles and abstracts in the four languages.
In order to have a multidirectional and trilingual balanced corpus, we have chosen
abstracts for which the original paper was written in English (five texts), Spanish
(five texts) and Basque (five texts). Thus, we have analyzed 15 abstracts (the same
ones for each language), written by different authors, constituting three subcorpora.
In sum, our corpus includes 45 texts. Table 2 summarizes the statistics of the
subcorpora.
In order to find correlations between translation strategies and rhetorical
relations, a methodology that can compare parallel rhetorical structures is needed.
We built our corpus in order to develop such a methodology, and consider that the
number of texts is sufficient for the design of the qualitative method that we present.
4
A problem with work in the framework of RST is that there is no annotated bilingual or trilingual
corpus to study the effects of translation strategies on rhetorical structure. As a consequence, a researcher
in such situation first needs to learn RST and perform annotations, as Maxwell (2010) suggests.
123
A qualitative comparison method for rhetorical structures
Table 2 Corpus statistics
Subcorpus
Annotators
Texts
Words
Sentences
EDUs
ENG
A1
15
5,706
201
318
SPA
A2
15
6,324
193
318
BSQ
A3
15
4,800
197
318
This qualitative method applies to any type of text,5 since the principles on which it
is based are general RST-based principles. We believe that the analysis is general
enough and the method applicable across genres. We also discuss some examples
detected with the qualitative evaluation in this parallel corpus that show how
translation strategies could be related to rhetorical structures (see Sect. 3.2.2).
After the corpus compilation, we carried out the analysis. This analysis had two
main phases: discourse segmentation and rhetorical analysis.
2.3 Discourse segmentation
The first step in analyzing texts with RST consists of segmenting the text into spans.
Exactly what a span is, in the framework of RST, and more generally in discourse, is
a well-debated topic. RST (Mann and Thompson 1988) proposes that spans, the
minimal units of discourse—later called Elementary Discourse Units (EDUs)
(Marcu 2000a)—are clauses, but that other definitions of units are possible.
From our point of view, adjunct clauses stand in clear rhetorical relations (cause,
condition, concession, etc.). Complement clauses, however, have a syntactic, but not
discourse, relation to their host clause. Complement clauses include, as Mann and
Thompson (1988) point out, subject and object clauses, and restrictive relative
clauses, but also embedded report complements, which are, strictly speaking, also
object clauses.
Other possibilities for segmentation exist; one of the better-known ones is the
proposal by Carlson et al. (2003) for segmentation of the RST Discourse Treebank
(Carlson et al. 2002). Carlson et al. (2003) propose a much more fine-grained
segmentation, where report complements, relative clauses and appositive elements
constitute their own EDUs.
In our work three annotators segmented the EDUs of each subcorpus (A1
segmented English texts, A2 segmented Spanish texts, and A3 segmented Basque
texts).6
5
It was used also to evaluate the RST Basque TreeBank (Iruskieta et al. 2013a), available at:
http://ixa2.si.ehu.es/diskurtsoa/en/.
6
When a corpus is annotated only with one annotator per language, the results may yield subjective
idiosyncrasies. This is not a problem for the aim of this paper, because we do not want to provide a
reliable annotated corpus in three languages, but we do provide a qualitative way to compare annotation
in different languages. Comparisons have been done manually and by pairs of languages following two
different evaluations: (a) Marcu’s quantitative method and (b) a new qualitative-quantitative method. So
even if the corpus is small, the comparison work is extensive. The aim to provide reliable corpora has
been achieved in other papers by the authors [English SFU corpus (Taboada and Renkema 2008), Spanish
RST TreeBank (da Cunha et al. 2011a) and Basque RST TreeBank (Iruskieta et al. 2013a)].
123
M. Iruskieta et al.
These annotators are experts in RST, having carried out research in this field for a
number of years, and they have participated in several projects related to the design
and elaboration of RST corpora in the three languages under consideration.
Annotators performed this segmentation task separately and without contact among
them. In our segmentation, we follow the general guidelines proposed by Mann and
Thompson (1988) which we have operationalized for this paper. We detail the
principles below.
2.3.1 Every EDU should have a verb
In general, EDUs should contain a (finite) verb. The main exception to this rule is
the case of titles, which are always EDUs, whether they contain a verb or not. Nonfinite verbs form their own EDUs only when introducing an adjunct clause (but not a
modifier clause; see ‘‘Appendix’’ for a detailed explanation).
2.3.2 Coordination and ellipsis
Coordinated clauses are separated into two segments, including cases where the
subject is elliptical in the second clause. In Spanish and Basque, both pro-drop
languages, this is in fact the default for both first and second clause, and therefore
we see no reason why a clause with a pro-drop subject cannot be an independent
unit. We follow the same principle for English.
Coordinated verb phrases (VPs) or verbs do not constitute their own EDUs. We
differentiate coordinated clauses from coordinated VPs because the former can be
independent clauses with the repetition of a subject; the latter, in the second part of
the coordination, typically contain elliptical verbal forms, most frequently a finite
verb or modal auxiliary.
2.3.3 Relative, modifying and appositive clauses
We do not consider that relative clauses (whether restrictive or non-restrictive),
clauses modifying a noun or adjective, or appositive clauses constitute their own
EDUs. We include them as part of the same segment together with the element that
they are modifying. This departs from RST practice, where (restrictive) relative
clauses are often independent spans, as seen in many of the examples in the original
literature and the analyzes on the RST web site (Mann and Thompson 1988; Mann
and Taboada 2010). We found that relative clauses and other modifiers often lead to
truncated EDUs, resulting in repeated use of the SAME-UNIT label,7 and thus decided
that it was best not to elevate them to the status of independent segments.
2.3.4 Parentheticals
The same principle applies to parentheticals and other units typographically marked
as separate from the main text (with parentheses or dashes). They do not form an
7
See the paragraph on Truncated EDUs in this section.
123
A qualitative comparison method for rhetorical structures
individual span if they modify a noun or adjective, but they do if they are
independent units, with a finite verb.
2.3.5 Reported speech
We believe that reported and quoted speech do not stand in rhetorical relations to
the reporting units that introduce them, and thus should not constitute separate
EDUs, also following clear arguments presented elsewhere (da Cunha and Iruskieta
2010; Stede 2008a). This is in contrast to the approach in the RST Discourse
Treebank (Carlson et al. 2003), where reported speech (there named ATTRIBUTION) is
considered as a separated EDU. There are, in any case, no examples of reported
speech in our corpus.
2.3.6 Truncated EDUs
In some cases, a unit contains a parenthetical or inserted unit, breaking it into two
separate parts, which do not have any particular rhetorical relation between each
other. In those cases, we make use of a non-relation label, Same-unit, proposed for
the RST Discourse Treebank (Carlson et al. 2003).
Once our segmentation criteria were established and the three annotators carried
out the segmentation, the three segmentations were compared in terms of F-measure
and Kappa. In this way, we quantified agreement and disagreement across
segmentations. Moreover, we analyzed the main causes of the disagreements. Results
are shown in Sect. 3.1. After the segmentation agreement evaluation, we harmonized
the segmentation, ensuring that units were comparable across the languages. At this
point, we also calculated linguistic distance between the pairs of languages, by
calculating which language required the most changes in the harmonization process.
This harmonization process was necessary to start out the analysis with similar units,
and to avoid confusing analysis disagreement and segmentation agreement. Marcu
et al. (2000) and Ghorbel et al. (2001) also align (which we termed harmonize) their
texts, decreasing the granularity of their segmentation to avoid complexity. With this
decision, we lose some rhetorical information at the most detailed level of the tree.
This does not, however, affect higher levels of tree structure. The results of this
harmonization are shown in Sect. 3.1.1.
2.4 Rhetorical analysis
Starting from the same discourse segmentation, we carried out the discourse
annotation of our corpus. Once again, A1 annotated English texts, A2 annotated
Spanish texts and A3 annotated Basque texts, using the mentioned extended
discourse relations set and RSTTool (O’Donnell 2000), a graphical interface widely
used for RST annotation. We compared the resulting rhetorical trees using two
different evaluation methods. One of them, which we characterize as a quantitative
evaluation, was proposed by Marcu (2000a), and the other one, which we describe
as a qualitative evaluation, was developed by our research team.
123
M. Iruskieta et al.
A qualitative comparison method for rhetorical structures in multilingual corpora
should quantify data, but also (and more importantly) should show linguistic
features affecting rhetorical structure. The quantitative/qualitative distinction is due
to the fact that the first method only gives us an approximate measure of agreement,
whereas the second method provides a qualitative description of annotation
dispersion. The qualitative evaluation, in addition to its use as a measure of interannotator agreement, can also be deployed to evaluate discourse structures built by a
parser.
2.4.1 Quantitative evaluation
In this section we present the quantitative method of Marcu (2000a) and its
limitations, already pointed out in other works (van der Vliet 2010; da Cunha and
Iruskieta 2010; Iruskieta et al. 2013b). The main limitations are:
1.
2.
Two of the factors evaluated, nuclearity and relation, are not independent of
each other: factor conflation.
The description of comparison and weight given to the agreement in certain
rhetorical relations could be improved: deficiencies in the description.
Marcu (2000a) presented a method to evaluate the correctness of discourse trees,
comparing automatically-built trees with manually-built ones. This method
measures recall and precision according to four factors: Elementary Discourse
Units (EDU), units linked with relations (Span), nuclear or satellite position
(Nuclearity) and rhetorical meaning of units (Relation). We refer to this method as
the quantitative method, because it uses exclusively numerical measures.
1. Factor conflation: nuclearity and relations. When measuring the relation
factor, the quantitative method conflates the label SPAN with a relation. Thus, the
SPAN label carries the same weight as any other relation. As we can see in Fig. 2, one
of the annotators has labelled the relation as ELABORATION, and the other as
EVIDENCE.
If we describe such disagreement with the quantitative method, we can see that
there is a degree of agreement with respect to the relation in the Fig. 3, when in fact
the agreement captured is simply the agreement in nuclearity, that is, in SPAN.
Figure 3 shows the results obtained after the comparison of the two rhetorical
structures included in Fig. 2 by using the quantitative evaluation. These results have
been obtained automatically by using RSTeval, which is an implementation of
Marcu’s comparison method.8
RSTeval does not take into account the language of the rhetorical structures;
however, it eliminates the stopwords of each language from the text, which are not
used to build the EDUs and Spans. In the first table of Fig. 3, absolute matches
between structures can be observed (e.g. Units: Matches = 2 of 2), as well as
percentages (e.g. Units: Recall = 1/Precision = 1), for the four mentioned factors.
8
This evaluation method has been automated by Maziero and Pardo (2009) and nowadays it can be used
in four languages: English, Spanish, Portuguese and Basque. Available at http://www.nilc.icmc.usp.br/
rsteval/.
123
A qualitative comparison method for rhetorical structures
Fig. 2 Quantitative evaluation: factor conflation (Iruskieta et al. 2013a, GMB0401)
The second table of Fig. 3 shows the detailed comparison process, where all the
constituents of the structures are included. In this case, the first constituent
corresponds to the first EDU, that is, words from ‘‘1 to 8’’ in the text; the second
constituent corresponds to the second EDU, that is, words from ‘‘9 to 13’’; and the
third constituent corresponds to the Span formed by the two mentioned EDUs, that
is, words from ‘‘!1 to 13’’ (the exclamation point at the beginning means that the
constituent is a Span). The symbol ‘‘x’’ indicates that a Unit or Span is included in
the corresponding rhetorical structure; ‘‘n’’ means nucleus; ‘‘s’’ means satellite, and
‘‘r’’ refers to the biggest span, that is, the span including the complete text. In the
Relations factor, if there is a nucleus, the category ‘‘span’’ is included when a
nuclear relation is under consideration or the name of relation when a multinuclear
relation is under consideration, while, if there is a satellite, the name of the
corresponding rhetorical relation is included.
Figure 4 shows a real example extracted from Iruskieta et al. (2013a).
In Table 3 we can see how RSTeval describes the agreement. The agreement
levels are shown in Table 4. For ease of reference, we have highlighted the
disagreements in italicize.
Fig. 3 Quantitative evaluation of Fig. 2 with RSTeval
123
M. Iruskieta et al.
Fig. 4 Annotations of text GMB0701 (Iruskieta et al. 2013a)
When examining the rhetorical relations factor, we can see that the SPAN label
plays a role in the description of agreement levels in Table 4: F-measure: 0.842 (16
agreements out of 19). If we describe the agreement without the SPAN label,
however, the degree of agreement changes, as we can see in Table 5: F-measure:
0.778 (7 agreements out of 9).9
9
Note that, after harmonizing discourse segmentation, accuracy, precision, recall and F-measure obtain
the same value. Therefore, although this results in a somewhat artificial level of agreement, we are
conscious about this fact, we use the standard measure employed in the RST literature (Marcu 2000a;
Maziero and Pardo 2009).
123
61 to 65 (prozesuaren_igurkapenen...dizkigute)
!51 to 65 (Horrez_item...dizkigute)
!36 to 65 (7_itemak...dizkigute)
!23 to 65 (Pierre_Martyren...dizkigute)
!16 to 65 (Basurtoko_Ospitaleko...dizkigute)
!5 to 65 (Ikerketa_Pierre...dizkigute)
!1 to 65 (Larritasunezko_irizpide...dizkigute)
!16 to 35 (Basurtoko_Ospitaleko...guztiei)
!5 to 35 (Ikerketa_Pierre...guztiei)
10
8–10
6–10
4–10
3–10
2–10
1–10
3–5
2–5
x
51 to 57 (Horrez_item...bereizten)
8
58 to 60 (horiei_balorazio...orokorra)
!36 to 50 (7_itemak...05)
6–7
!51 to 60 (Horrez_item...orokorra)
39 to 50 (estatistikoki_desberdintasun...05)
7
9
36 to 38 (7_itemak...aztertuta)
6
8–9
x
!23 to 35 (Pierre_Martyren...guztiei)
4–5
x
x
x
x
x
23 to 31 (Pierre_Martyren...asmoz)
32 to 35 (elkarrizketa_zitzaien...guztiei)
x
4
16 to 22 (Basurtoko_Ospitaleko...gaixok)
3
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
A3
A3
A4
Spans
Units
5
1 to 4 (Larritasunezko_irizpide...onkologian)
5 to 15 (Ikerketa_Pierre...aztertu)
1
2
Constituent
EDU
Table 3 Qualitative method for text GMB0701
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
A4
n
s
r
n
s
n
n
n
n
n
n
n
s
s
n
s
n
n
s
A3
N/S
r
n
s
s
s
n
n
n
n
n
n
n
s
n
n
s
n
n
s
A4
Span
Means
Span
Span
Result
List
List
List
List
List
List
Span
Means
Elaboration
Span
Purpose
Span
Span
Preparation
A3
Relations
Span
Span
Means
Elaboration
Result
List
List
List
List
List
List
Span
Means
Span
Span
Purpose
Span
Span
Preparation
A4
A qualitative comparison method for rhetorical structures
123
M. Iruskieta et al.
Table 4 Quantitative method: agreement level for text GMB0701
Units
Spans
N–S
Relations
Match
R
P
Match
R
P
Match
R
P
Match
R
P
10 of 10
1
1
17 of 19
0.895
0.895
16 of 19
0.842
0.842
16 of 19
0.842
0.842
Table 5 Agreement level according to rhetorical relations in GMB0701
Relations
Match
R
P
7 of 9
0.778
0.778
2. Deficiencies in the description. When annotators decide that a relation has an
attachment point at different levels in the tree structure (da Cunha and Iruskieta
2010), the method proposed by Marcu (2000a) is not able to compare the relations
where constituents has changed. Observe the following issues in Fig. 4:
–
–
In Table 3 the agreement in the ELABORATION relation cannot be included,
because the relation has different spans: in A3 ‘23 to 31’ and in A4 ‘!23 to 65’
both attachments are referred as the same constituent, ‘23 to 31’.
The MEANS constituent of A3 ‘!16 to 35’ and in A4 of ‘!16 to 65’, both attach to
the same EDU (EDU2 or ‘5 to 15’); but, since the constituents do not coincide,
the two MEANS relations cannot be compared.
Following da Cunha and Iruskieta (2010), Iruskieta et al. (2013b) and Mitocariu
et al. (2013), we think that a qualitative method should describe the six factors
involved in all rhetorical relations independently: EDU and Span (segmentation),
nucleus-satellite function (Nuclearity), and attachment point, constituent and
rhetorical meaning (Relation). When parallel texts are compared, a qualitative
method should take in account whether the language form is parallel, as explained in
the next section.
2.4.2 Qualitative evaluation
The qualitative evaluation method that we propose considers both type of agreement
and source of disagreement, which results in a better explanation of the dispersion in
annotator interpretations about text structure. When analyzing rhetorical structures
using Marcu’s method, we observed that similar structures at the intermediate level
of a tree structure spans could not be compared, because the constituents did not
coincide. Such structures had, however, the same rhetorical relation, and the fact
that the relation is the same should be reflected in a measure of agreement. If we
accept that constituents do not need to coincide in their (span size) entirety to be
compared, the issue is whether we can state that there is agreement with respect to
the rhetorical relation, but disagreement about the constituents.
123
A qualitative comparison method for rhetorical structures
In our evaluation method it is not necessary for the constituents to be compared
to be identical, like in Marcu’s (2000b) method; only the central subconstituent (CS)
has to be the same.10 With such restriction we are able to compare rhetorical
relations, using four independent criteria: constituent, attachment point, the
direction of the relation (nuclearity) and effect of the relation.
When comparing RST structures with independent factors, we do not use typical
nucleus and satellite terms to describe the extension of spans, because our method
assesses independently nuclearity and unit size. The comparison in our method is
based on rhetorical relations and not in spans of relations as Marcu’s (2000b)
method does. In our method we have a line for each relation, while in Marcu’s
(2000b) method there are two lines for each relation. The term constituent (C) refers
to the length of the constituents, and the term attachment point (A) refers at the
height of the tree where the constituent is linked (in Marcu’s (2000b) evaluation
method this factor is not considered, because what is compared are spans of
relations). Because we are comparing relations and not spans of relations, in our
comparison also nuclearity has a different meaning; while in Marcu’s (2000b)
method nuclearity has two possible values (S or N, where S means satellite and N
means nucleus) for each span, in our method nuclearity has three values (SN, NN
and NS) for each relation.
First of all, we present the types of agreement, and the two sources of
disagreement in the qualitative evaluation by comparing annotators’ RST trees. We
measure the agreement in rhetorical relations based on the following factors:
constituent (C), attachment point (A) and the name of relation (R), checking some
agreement types:
1.
2.
3.
4.
Agreement
Agreement
Agreement
Agreement
in relation, constituent and attachment point (RCA).
in relation and constituent (RC).
in relation and attachment point (RA).
only in relation (R).
A decision tree formalizes the method to check the agreement types in rhetorical
relations (see Fig. 5). As we mentioned before, to check agreement in rhetorical
relation, the constituent of this relation must have the same central subconstituent
(CS). If this condition is fulfilled, we check if relation name (R), constituent (C) and
attachment point (A) are exactly the same.
We distinguish two sources of disagreement, disagreements of type A and type L,
for Annotator and Language disagreements:
Disagreements of type A (Annotator). No significant linguistic differences in the
text, but distinct relations labelled by two annotators (marked with an ½A in column
Disagree of Table 7, and in corpus results in Table 17 under Annotation
Discrepancies). We have found seven sources of such disagreement:
1.
2.
Different choice in nuclearity entailed a N/N–N/S mix-up (N/N–N/S).
Different choice in nuclearity entailed discrepancy in N/S relations (N/S).
10
If there is more than one CS (because there is a multinuclear relation) at least one of them has to be the
same for N/S-N/N mix-up.
123
M. Iruskieta et al.
Fig. 5 Decision tree based on CS to establish the agreement types about R
3.
4.
5.
6.
7.
A relation has the same constituent and attachment point, but not the same
relation label (6¼ R).
Relations chosen are similar in nature (Similar R).
Relations with mismatched RST trees (Mismatch R).
A relation is more specific than the other (Specificity).
Different choice in attachment entailed a different relation (Attachment).
Disagreements of type L (Language). Two annotators labelled distinct relations
because there is a significant difference in the linguistic form (marked with an ½L in
column Disagree of Table 7 and in corpus results in Table 20 under Translation
Strategies). We have found three different sources. These are in fact translation
strategies, and are sensitive to corpus and language. Studies in other corpora, genre
or languages may reveal different strategies and sources of disagreement:
1.
2.
3.
A relation is signaled with a different discourse marker (Marker Change or
MC).
A different organization of constituent phrases is used, mostly from non-finite
verb phrase to finite verb phrase (Clause Structure Change or CSC).
A change in unit level (phraseclausesentence) is done (Unit Shift or US).
In Table 6 we show an example extracted from the corpus of text TERM38_SPA
which was segmented and harmonized in Spanish (A2) and in English (A1) (Fig. 7)
to illustrate the qualitative method (Table 7).11
11
Basque segments (A3) were also harmonized, but space constraints preclude us to align with Spanish
and English. Anyway, the harmonization of TERM38_SPA segmentation in the three languages can be
consulted at: http://ixa2.si.ehu.es/rst/segmentuak_multiling.php?bilatzekoa=TERM38%. The English RStree can be consulted at: http://ixa2.si.ehu.es/rst/diskurtsoa_jpg/TERM38_A1.jpg. The Spanish RS-tree
can be consulted at: http://ixa2.si.ehu.es/rst/diskurtsoa_jpg/TERM38_A2.jpg.
123
A qualitative comparison method for rhetorical structures
Fig. 6 Decision tree to establish the sources of agreement and disagreement about R
Table 7 includes the analyzed factors for Fig. 7: nuclearity (N), relation (R),
constituent (C) and attachment point (A). These factors compare A2 (Spanish) and
A1 (English). In the Qualitative Evaluation columns, we mark with a ‘‘U’’ an
instance of agreement, and with an ‘‘’’ a disagreement. The last two columns
summarize the type of agreement (Agree) or the disagreement source (Disagree).
If there is a multinuclear relation inside of a constituent of another relation (see
lines 22 and 23 in Table 7) comparing CSs is not trivial, because multinuclear
relations have more than one CS. Line 23 is representative of this problem. If we
look at this line we can see that the problem is not the relation that we are
comparing, but the problem comes from a lower level, since there is full agreement
(RCA) between annotators (on R: ELABORATION, on C: 11N and on A: 12–14S).
When this is the case there are two choices: (a) do not compare relations and
annotate as ‘‘no-match’’12 and (b) compare first non-ambiguous CSs and leave
problematic comparisons (lines 22 and 23) for the end. Following the last choice
there is not any ambiguous CS in Table 7, because the other CS candidate (CS 12 in
line 10) was used in other structure. Because of that, when we have to compare
relations with more than one CS with another that has only one CS, at least one of
the CSs has to be identical. If still there were cases in which we can not compare
structures we have used the no-match label. This problem was found also in text
summarization by Marcu Marcu (2000b), since the most important unit can be
formed by more than one EDU.13
In Table 8 we present the results of our evaluation method for the example in
Fig. 7.
12
If we follow this decision, we could not compare structures that contain a N/N–N/S mix-up inside the
relation.
13
As the evaluation has been done manually, there have been some problematic cases that have not
counted as an agreement. For cases in which some structures cannot be compared, no-match label has
been used, which represents not more than 0.06 % of all relations (53 no-match/900 relations), about 1.18
relations per text on average (53 No Match/45 texts).
123
123
Spanish
9
1 to 6
7 to 22
23 to 38
39 to 67
68 to 92
93 to 105
106 to 123
124 to 164
165 to 173
7
1
2
3
4
5
6
7
8
9
Es más, desde cualquier lugar los términos son recopilados,
comentados y ponderados
En primer lugar, el canal por el que se dan a conocer los términos
de Internet, la misma red, no sólo supone una rápida difusión de
la terminologı́a—la información en Internet es de acceso (casi)
inmediato—, sino también un alcance muy vasto—llega a
cualquier parte del mundo—
la especificidad del área tratada confiere a la neologı́a que le es
propia unas particularidades que cabe tener en cuenta
Si bien este aspecto es común al progreso cientı́fico y técnico y, por
lo tanto, caracterı́stico de la neologı́a terminológica
Efectivamente, la formación de nuevos términos está sometida a un
ritmo trepidante, paralelo al avance e innovación tecnológica en
el sector de la informática y, en general, de las
telecomunicaciones
Los términos referidos a Internet nacen y se difunden a una
velocidad y con una amplitud tal que constituye una verdadera
carrera contrarreloj en las distintas lenguas
para lo cual vamos a abordar diversos aspectos que influyen en la
creación neológica en el ámbito de Internet
El propósito de esta comunicación es hacer una reflexión sobre los
retos a que se está enfrentando la neologı́a terminológica en la
realidad actual
La neologı́a contrarreloj: Internet
Languages
Tables
Table 6 TERM38_SPA segmented and harmonized in Spanish and English
Furthermore, terms can be compiled, discussed and assessed anywhere
First of all the channel through which Internet terms are made known is the
net itself. This means that they not only spread rapidly (information on
the internet can be accessed almost immediately) but also reach vast
areas (all over the world)
but the specific nature of this area confers particular features on neology
which must be taken into account
This is common in all scientific and technological progress, and therefore
characteristic of neology in terminology
The formation of new terms goes on at a dizzy speed, parallel to
technological advances and innovations in the field of computer science
and telecommunications in general
Terms referring to the Internet are coined and spread at such speed and to
such an extent that they have turned into a race against the clock in
different languages
I will do this by discussing various points which influence neology in the
field of the Internet
This paper is intended to look at the challenges faced by neology in
terminology at the present time
Neology against the clock: the Internet
English
M. Iruskieta et al.
Spanish
9
174 to 196
197 to 203
204 to 224
225 to 229
230 to 256
257 to 262
263 to 267
268 to 273
274 to 278
278 to 281
282 to 289
290 to 296
297 to 307
308 to 314
315 to 320
7
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
>Qué sistemas de creación léxica predominan?
Are all those words which seem to be terms actually terms?
>Son términos todos los que lo parecen
o abundan las creaciones léxicas sensacionalistas y efı́meras?
or do sensationalist, ephemeral terms abound?
Do they meet actual needs for names
>Cómo tratan la terminologı́a de Internet?
responden a necesidades reales de denominación
How do the receiving languages respond to this?
How do they deal with Internet terminology?
>Cómo responden las lenguas receptoras?
and come in as loanwords
terms are generated in English
y penetran como préstamos en aquellas
los términos se generan en inglés
There is a common denominator in all languages
What type of terminology is being created?
What lexical creation systems predominate?
>Qué tipo de terminologı́a se está creando?
Un único denominador común existe para todas las lenguas
It is used both by a wide variety of net users (from any or no specialist
fields) and by people who read the press or follow the media
and breaks into general language
Internet terminology extends beyond the bounds of its specialist field
(which by definition is part of the lexicon of science and technology)
This leads us to the fundamental point
many Web sites can be found which give glossaries of Internet terms or
propose names and even invite users to vote on them
English
siendo utilizada tanto por los usuarios heterogéneos de la red (de
cualquier o ninguna especialidad) como por las personas que
leen la prensa o están atentas a los medios de comunicación
e irrumpe en la lengua de uso general
la terminologı́a de Internet traspasa los lı́mites del área de
especialidad (a la que se circunscribe por definición el léxico
cientı́fico y técnico)
Esto nos lleva a una cuestión fundamental
de ahı́, por ejemplo, los apartados que encontramos en muchos
Webs en que se difunden glosarios de términos sobre Internet o
en que se exponen propuestas denominativas que los usuarios
pueden incluso votar
Languages
Tables
Table 6 continued
A qualitative comparison method for rhetorical structures
123
M. Iruskieta et al.
Fig. 7 Rhetorical tree elaborated by A2 (Spanish) and A1 (English), TERM38_SPA
In order to better highlight the differences between the quantitative method and
our qualitative proposal, we have kept the rhetorical structure, but have used one of
the languages to compare using RSTeval in contingency Table 9.
123
15|16–24
15|16|20–24
16|20|21|22–24
17
12
13
14
15
22–23|24
8
12|13
21
22
23
22|23–24
14
11
20
12
10
21|22–24
11
9
20|21–24
10
8
19
9
7
18
6
6
18–19
7
5
18|19
5
4
16
4
3
17
1
3
1
Preparation!
Elaboration
Elaboration
Contrast$
List$
List$
List$
Sequence$
Elaboration
Elaboration
List$
Elaboration
List$
Result
List$
Interpretation
Evidence
Elaboration
Concession!
Elaboration
Elaboration
Elaboration
Means
1S
12–14S
8–24S
22–23N
22N
21N
20N
18N
18–19S
17–19S
20–24N
15–24S
15N
14S
12N
11–14S
10S
9–10S
6S
6–7S
5–7S
4–24S
3S
11N
4–7N
24N
23N
22–24N
21–24N
19N
17N
16N
16–19N
8–14N
16–24N
13N
13–14N
8–10N
9N
8N
7N
5N
4N
2–3N
2N
2–24N
A
13
8|9
23|24
22|23–24
21|22–24
20|21–24
18|19
18–19
17
20|21|22–24
15|16|20–24
15|16–24
14
12
11
10
8|9
6
7
5
4
3
1
CS(s)
C
CS(s)
R
SPA
ENG
2
L
Table 7 Qualitative evaluation matrix TERM38_SPA
Elaboration
Elaboration
Disjunction$
List$
List$
List$
Sequence$
Elaboration
Background!
Elaboration
Elaboration
List$
Elaboration
Cause
Background!
Elaboration
List$
Concesion!
Elaboration
Elaboration
Elaboration
Background!
Preparation!
R
12–14S
8–10S
23N
22N
21N
20N
18N
18–19S
17–19S
17–24S
3–24S
15N
14S
12S
11–14S
10S
9–10N
6S
6–10S
5S
4–10S
3–10S
1S
C
11N
6–7N
24N
23–24N
22–24N
21–24N
19N
17N
20–24N
16N
2N
16–24N
13N
13–14N
15–24N
9N
8N
7N
4–5N
4N
3N
11–24N
2–24N
A
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
U
A
U
C
U
U
U
U
U
U
U
U
U
U
U
R
N
Qualitative evaluation
RCA
R
ðCÞ
RCA
RCA
RCA
RCA
RCA
ðAÞ
R
RCA
ðCAÞ
ðCAÞ
ðAÞ
ðCAÞ
ðCAÞ
RCA
R
RA
R
RCA
Agree
6¼ R½A
N=S½A
N=S½A
CSC½L
N=NversusN=S½A
N=S½A
MC½L
N=NversusN=S½A
N=S½A
Disagree
A qualitative comparison method for rhetorical structures
123
M. Iruskieta et al.
Table 8 Qualitative evaluation results for the example in Fig. 7, TERM38_SPA
Nuclearity
Relation
Composition
Attachment
Matches
F1
Matches
F1
Matches
F1
Matches
F1
16 of 23
0.6957
14 of 23
0.6087
15 of 23
0.6522
16 of 23
0.6957
Both methods measure the similar factors: (1) EDUs and spans (constituent and
attachment), (2) nuclearity (of each unit, or direction of the relation) and rhetorical
relations (of each unit: relation plus span, or relation as a whole). Thus, in Table 11
we can compare how each method accounts for these factors.
In Table 11 both methods describe total agreement in segmentation. This is due
to the fact that segmentation was harmonized before the analysis was undertaken.
The span factor of the quantitative method is described using factors C and A, this
factor being more positive in the quantitative method. In terms of nuclearity and
rhetorical relations, the qualitative method is able to describe more agreements in
the evaluation of text TERM38.
In Table 12 we can observe further detail on how both methods describe
agreement in relations, and the weight given to each relation in the calculation of
agreement. To better understand the table, we have highlighted in italicize the most
important differences.
As we can see in Table 12, an important part of the agreement in quantitative
evaluation method is captured in the SPAN label (which is not an RST relation). In
addition, the contingency table shows that the relation with most agreement is the
LIST relation, followed by ELABORATION and SEQUENCE. Thanks to the qualitative
evaluation, however, we can see that the ELABORATION relation actually has a higher
degree of agreement, followed by LIST. In contrast, SEQUENCE has little importance,
the same as CONCESSION and PREPARATION. We would like to point out that the
difference is more striking when describing agreement (Match: columns 4 and 8),
rather than when describing how often the annotator has used such relation (A1:
columns 2 and 6, and A2: columns 3 and 7). For instance, in both methods we can
see that A1 has used 10 ELABORATION relations, whereas A2 has used 9 relations. The
quantitative method captures an agreement of 4.35 %, while the qualitative method
throws a much higher agreement, reaching 26.09 %.
The root of this difference can be found in the fact that the quantitative
evaluation does not evaluate nuclearity and rhetorical relations in an independent
way. When creating relation pairs, the pairs do not have well-formed members (in
particular because of the use of the SPAN label). This is the reason why in the
quantitative method, out of 10 ELABORATION relations, only two of them show
agreement.
Advantages of the qualitative evaluation method. The formalization of qualitative
evaluation (Table 7) describes the annotation agreement (Agree) in a more complete
way than quantitative evaluation (Table 9): the relation factor (R) is compared in an
isolated manner, that is, nuclearity is not reanalyzed in the relation factor. This fact
has methodological implications and some of advantages are shown in contingency
Table 7:
123
x
x
x
174 to 196
x
x
x
204 to 224
225 to 229
230 to 256
x
x
x
!204 to 256
!197 to 256
x
x
!225 to 256
x
x
x
x
x
x
197 to 203
x
x
!124 to 196
x
x
!165 to 196
x
x
x
x
x
x
x
x
x
x
x
x
x
165 to 173
x
x
124 to 164
x
x
x
x
!39 to 123
x
x
x
x
x
x
x
!68 to 123
!93 to 123
x
x
106 to 123
x
x
x
93 to 105
x
x
68 to 92
x
x
x
x
x
x
x
x
x
x
x
39 to 67
x
23 to 38
x
x
x
x
A2
!7 to 38
x
7 to 22
A1
A1
A2
Spans
Units
1 to 6
Constituent
s
s
n
s
n
s
n
s
n
s
n
n
n
n
s
s
n
n
n
s
A1
s
s
n
s
n
n
n
n
s
s
n
n
n
s
s
n
s
n
n
n
s
n
s
A2
Nuclearity
Background
Elaboration
Span
Elaboration
Span
Cause
Span
Elaboration
List
Elaboration
Span
List
Span
Span
Concession
Elaboration
Span
Span
Span
Preparation
A1
Relation
Interpretation
Elaboration
List
result
Span
List
Span
Span
Elaboration
Evidence
Span
Span
Span
Elaboration
Elaboration
Span
Concession
Span
Span
Span
means
Span
Preparation
A2
!282 to 296
!23 to 196
!39 to 196
!93 to 196
!39 to 92
!1 to 320
!7 to 320
!39 to 320
!124 to 320
!263 to 320
!257 to 320
!297 to 320
315 to 320
!297 to 314
308 to 314
297 to 307
!257 to 296
290 to 296
!257 to 289
282 to 289
257 to 262
!263 to 281
!268 to 281
Constituent
Table 9 Contingency table for text TERM38_SPA with quantitative method, using RSTeval
x
x
x
x
x
x
A1
Units
x
x
x
x
x
x
A2
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
A1
Spans
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
x
A2
n
s
s
s
n
r
n
n
n
n
n
n
n
n
n
s
A1
r
n
s
s
s
n
n
n
n
n
n
n
n
n
n
n
n
s
A2
Nuclearity
List
background
Elaboration
Elaboration
Span
Span
Span
List
Span
Disjunction
Disjunction
List
List
List
List
Background
A1
Relation
Span
Span
Elaboration
Elaboration
Elaboration
List
List
Contrast
Contrast
List
List
List
List
List
List
List
List
Elaboration
A2
A qualitative comparison method for rhetorical structures
123
123
x
278 to 281
!274 to 281
x
x
x
x
x
x
x
x
x
x
274 to 277
x
x
268 to 273
x
x
x
x
x
x
A2
263 to 267
A1
A1
A2
Spans
Units
!124 to 256
Constituent
Table 9 continued
s
n
n
n
n
A1
s
n
n
n
n
n
A2
Nuclearity
Elaboration
Sequence
Sequence
Span
Span
A1
Relation
Elaboration
Sequence
Sequence
Span
Span
Span
A2
!23 to 320
!197 to 320
!268 to 320
!282 to 320
!308 to 320
!282 to 307
Constituent
A1
Units
A2
x
x
x
x
x
x
A1
Spans
A2
s
n
s
n
n
n
A1
A2
Nuclearity
Elaboration
Span
Elaboration
Span
List
List
A1
Relation
A2
M. Iruskieta et al.
A qualitative comparison method for rhetorical structures
Table 10 Quantitative method results for text TERM38_SPA
Units
Span
Nuclearity
Relation
Match
F1
Match
F1
Match
F1
Match
F1
24 of 24
1
36 of 47
0.766
29 of 47
0.617
20 of 47
0.425
Table 11 Comparison using both methods, TERM38_SPA
Units
Quanti.
24 of 24
Spans
1
Units
Quali.
24 of 24
1
Nuclearity
37 of 46
0.8043
29 of 46
Composition
Attachment
Nuclearity
15 of 23
14 of 23
17 of 23
0.6522
0.6087
Relation
0.6304
21 of 46
0.4565
Relation
0.7391
13 of 23
0.5652
Table 12 Comparison of agreement using both methods for text TERM38
Relation
Quantitative method
A1
A2
Qualitative method
Match
%
A1
Background
3
3
Cause
1
1
Concession
1
Contrast
Disjunction
2
Elaboration
10
Evidence
Means
1
1
Match
%
1
1
4,35
6
26,09
4
17,39
1
4,35
1
1
9
2
4.35
10
12
1
9
1
1
6
13.04
5
1
Result
Sequence
2.17
1
10
Preparation
1
1
Interpretation
List
1
2
A2
6
1
1
2.17
1
1
1
1
2
2
2
4.35
1
1
1
4,35
Span
16
15
9
19.57
–
Total
46
46
21
45.65
23
23
13
56,52
1.
Independent factors are evaluated. A different attachment point of a relation
only implies disagreement in attachment point (disagreement described at the
same line) and in constituent (disagreement described at a higher level in the
tree structure) and not in relation as quantitative method does. Moreover, the
qualitative method accounts for the source of disagreement (Disagree).
123
M. Iruskieta et al.
2.
3.
4.
Only rhetorical relations are compared. The description allows for a full
coincidence in structure (RCA), or a partial match (RA, RC or R).
Reasons for annotator disagreement are captured: aÞ because of differences in
the linguistic expression ½L or bÞ because of interpretation ½A.
Relation pairs in the contingency table are able to better describe agreement and
disagreement (‘‘confusion patterns’’, Marcu 2000a).
For example, in Table 7 we can observe the following types of information on the
relation agreement:
1.
2.
3.
Match in relation, constituent and attachment point (RCA) in the following nine
lines: 1, 6, 12, 16, 17, 18, 19, 20 and 23. We observe that in these lines there
was total agreement in the three factors observed, that is, for example, in line 1
an agreement in all factors: same CS (1), relation (PREPARATION), constituent
(1S) and attachment point (2–24N).
Match in relation and attachment point (RA) in line 4. A partial agreement, but
in this case in CS (5), relation (ELABORATION) and attachment point (4N). By
contrast, slight disagreement in constituent (A2: 5–7S but A1: 5S).
Match only in relation (R) in four lines: 3, 5, 13 and 22. For example, in line 3
there was an agreement only in CS (4) and relation (ELABORATION), whereas
there were discrepancies in constituent (A2: 4–24S but A1: 4–10S) and
attachment point (A2: 2–3N but A1: 3N).
On the relation disagreement, we can observe the following types of information in
Table 7:
1.
2.
3.
4.
5.
A
A
A
A
A
different choice in nuclearity (N/S [A]) in four lines: 2, 9, 14 and 15.
N/N–N/S mix-up (N/N–N/S [A]) in two lines: 7 and 10.
different relation label (6¼ R [A]) in a line: 21.
Marker Change (MC [L]) in a line: 8.
Clause Structure Change (CSC [L]) in a line: 11.
3 Results
In this section, we first present the results of segmentation, and then we compare the
results of rhetorical structure based on two evaluation methods: quantitative method
(Marcu 2000a) and our new proposal, a qualitative evaluation method.
3.1 Discourse segmentation results
The initial round of segmentation led to the following number of EDUs: 330 in
English, 318 in Spanish, and 323 in Basque. We calculated agreement using F-score
and Kappa, in a pairwise manner. First of all, we calculated the total coincidence of
EDUs, using the verb of the main clause and its principal arguments (VP). If the
main verb was the same in both EDUs, then we tabulated it as a match. As we stated
in page 7, one of our segmentation principles is that every EDU should contain a
123
A qualitative comparison method for rhetorical structures
Table 13 Segmentation agreement
Language
Correct
Match
Wrong
Missing
Candidates
F-measure
Kappa
ENG-SPA
330
230
88
12
731.4
70.99
0.7139
ENG-BSQ
330
226
97
7
742.9
69.22
0.7057
BSQ-SPA
323
230
88
5
731.4
71.76
0.7333
finite verb. The main verb of an EDU indicates the principal action, process, state,
condition, etc., in relation to the subject of the clause. Therefore, if two EDUs in
different languages contain the same verb (that is, both verbs are translation
equivalents), they are expressing the same event and we consider that there is
coincidence between EDUs. Thus, in this sense, syntax has an important role to play
in the detection of the EDUs to be compared, since we take the main verb of the
clausal syntactic structure in each language to carry out the comparison. In this
work, we have not used a syntactic parser to perform the analysis. We have done the
analysis manually, because it was feasible to do it over our corpus and we also
wanted to avoid possible mistakes in the harmonization work.14 In future work,
however, we plan to automate our methodology to compare discourse structures,
and, in this case, we could integrate a syntactic parser in the system. We then
calculated F-measure and Kappa as presented in Table 13.15
3.1.1 Discourse segmentation harmonization
In our segmentation, it was often the case that one language used a finite verb,
whereas the other language used a non-finite verb or other expression, leading to
differences in segmentation. Another source of disagreement was the interpretation
of ellipsis, where one annotator decided there was more than subject ellipsis in
coordination, and did not break up the two VPs, whereas the other annotator decided
to break them up. Two other sources of disagreement were different texts in the two
languages (not different formulations, but a completely different text, with one
sentence deleted or inserted), and simple human error. The latter accounts for no
more than two disagreements per language pair.
Harmonization led to joining or separating EDUs in one of the languages,
contravening our general principles for segmentation. The main changes in this
harmonization were:
1.
2.
When two parallel passages share the same structure and the third passage does
not, then we harmonize the segmentation of the third language taking into
account the segmentation of the two coincident languages.
When the segmentations of the three parallel passages are different, then we
harmonize the segmentation taking into account the structure of the simplest
passage.
14
This harmonization work can be found at http://ixa2.si.ehu.es/rst/segmentuak_multiling.php.
15
For Kappa segment candidates were calculated automatically by counting verbs.
123
M. Iruskieta et al.
In Example (1) a Basque conjunct was translated as a clause in both English and
Spanish. In the English example there are three finite verbs (all three of them
instances of the verb is), as is the case in Spanish (es, ‘½it is’; se ubica, ‘½it is
located’; and va, ‘½it goes’). In Basque, however, there are only two finite verbs
(estrapolatuko du, ‘½it will extrapolate ½it’; and jartzen du, ‘½it places ½it’). The
third part of the conjunct contains no verb (eta hizkuntza erromanikoek ezkerraldean, ‘and the Romance languages on the left side’). In the harmonization we
inserted a new segment in Basque, reinterpreting not as coordinated NP, but as a
juxtaposed clause with an elided verb.16
(1)
(a)
[Our hypothesis is that a syntactic characteristic of Basque and the
romance languages is extrapolated to their morphology,] [so that in
Basque derivations the core of the structure is on the right,] [while in the
romance languages it is on the left.]
(b) [Nuestra hipótesis es que una caracterı́stica sintáctica del euskera y de las
lenguas románicas se extrapola hasta la morfologı́a,] [de manera que en
euskera, también en derivación, el núcleo de la estructura se ubica a la
derecha,] [mientras que en las lenguas románicas va a la izquierda.]
(c) [Gure hipotesiak, euskararen eta hizkuntza erromanikoen ezaugarri
sintaktiko bat morfologiaraino estrapolatuko du:] [eratorpenean ere
euskarak egituraren burua edo gunean eskuinaldean jartzen du,} {eta
hizkuntza erromanikoek ezkerraldean.] TERM50_BSQ
In Example (2) the translation from Spanish into English has led to two separate clauses.
The Spanish original segmentation contained only one span, since the first idea (un
aumento cuantitativo de la terminología especializada, ‘an increase in the number of
specialist terms’) is embedded in a non-finite clause (además de provocar, ‘in addition
to leading to’). The English translation splits the ideas into two coordinated clauses
(factors lead to an increase and but also [factors] call into question). Basque also has
two clauses to express these two ideas. Since two of the languages divided this sentence
into two clauses, in the harmonization we inserted a new boundary in Spanish.
(2)
(a)
[All these factors lead to an increase in the number of specialist terms which
enrich terminology] [but also call into question some of its basic concepts,
such as the one to one relationship between ideas and names, the concept of
mastery of a specialist field and the role of standardization in terminology.]
(b) [Todos estos factores, además de provocar un aumento cuantitativo de la
terminologı́a especializada, han implicado una ampliación de la
perspectiva del trabajo en terminologı́a,} {que si bien la ha enriquecido,
al mismo tiempo ha puesto en cuestión algunos de sus conceptos básicos,
como la univocidad noción-denominación, el concepto de dominio de
especialidad o el papel mismo de la normalización en terminologı́a.]
16
In the example, the original segmentation is marked with square brackets and the segmentation after
harmonization with curly brackets.
123
A qualitative comparison method for rhetorical structures
(c)
[Alderdi horiek guztiek, espezialitateko terminologiaren gehikuntza
kuantitatiboa eragiteaz gain, terminologia lanen ikuspegia ere zabaldu
egin dute;] [eta, egia bada ere ikuspegi berri horrek terminologia aberastu
egin duela esatea, zalantzan jarri ditu terminologiaren oinarrizko zenbait
kontzeptu: kontzeptu-izendapen bikotearen adierabakartasuna, espezialitateko eremuen kontzeptua, eta normalizazioak terminologian duen
eginbeharra.] TERM19_SPA
We quantified the changes necessary to harmonize the segmentations by counting
how many times a change was necessary, per language. Table 14 summarizes those
changes (the typical actions are ‘‘join’’ or ‘‘break up’’), and the number of affected
EDUs. To compute the number of affected EDUs, we counted, in the cases where
we needed to break down a unit, how many new units were necessary (þ). In the
cases where we needed to join, we counted how many original units were integrated
(). In the table, ‘‘initial spans’’ refers to the spans proposed by the individual
annotator for each language, and ‘‘affected spans’’, to the number of spans that
underwent a change, whether to join, or to break up. ‘‘Harmonized spans’’
represents the final agreed upon spans across all three languages, for each text.
We can see from the table that the language with more changes is Basque.17 We
found that the linguistic expression of the same or similar concepts required
different syntactic constructions in Basque. This makes sense, given that Basque is a
non-Indo-European language, showing considerable typological distance from both
Spanish and English (Cenoz 2003). Note that, whereas Spanish and Basque were
affected in the same proportion in both directions (when breaking down SPA:
44.44 % and BSQ: 41.46 %; when joining SPA: 55.56 % and BSQ: 58.54 %),
harmonization in English involved breaking down in a much lower proportion
(when breaking down ENG: 18.18 %; when joining ENG: 81.82 %). This suggest
that the corpus abstracts in English (whether translated or original) express clauses
as separate units, either as simple sentences or as clear (finite) adjunct clauses,
without using non-finite clauses or prepositional complements.
3.2 Rhetorical analysis results
Results of quantitative method were presented in order to show the consistency of
this method. To this end, first, we present below the results of the quantitative
method; second, we present the results of the qualitative method, and after that we
compare results from both methods.
3.2.1 Results of the quantitative evaluation method
Results of the quantitative evaluation are shown in Table 15.18
17
One-way ANOVA demonstrated significant differences across the three languages in the corpus
(p ¼ 0:07). We thought this was quite significant, therefore we performed a post-hoc Tukey’s test and we
observed that harmonization in Basque is the furthest from the other two.
18
EDUs are excluded because they are identical after harmonization.
123
M. Iruskieta et al.
Table 14 Segmentation changes
Initial spans
Text
Harmon.
Affected spans
Spans
ENG
SPA
BSQ
ENG
SPA
BSQ
TERM18_ENG
8
11
14
8
0
3
6
TERM19_SPA
14
12
13
14
0
þ2
þ1
TERM23_ENG
15
14
14
14
1
0
0
TERM25_BSQ
10
11
8
10
0
þ1
þ2
TERM28_BSQ
16
14
12
15
1
þ1
þ3
TERM29_SPA
14
14
13
14
0
0
þ1
TERM30_ENG
26
27
33
28
þ2
þ1
5
TERM31_BSQ
53
52
44
52
1
0
þ8
TERM32_ENG
13
13
18
13
0
0
5
TERM34_BSQ
50
45
44
46
4
þ1
þ2
TERM38_SPA
27
25
28
24
3
1
4
0
TERM39_ENG
7
8
9
9
þ2
þ1
TERM40_SPA
8
8
8
8
0
0
0
TERM50_BSQ
34
35
30
30
4
5
0
4
35
29
35
31
4
þ2
330
318
323
316
22
18
41
6.67 %
5.66 %
12.69 %
TERM51_SPA
Total
Change rate
Table 15 Quantitative evaluation results (F-measure)
Language comparison
Evaluation
1st Lang.
2nd Lang.
Span (%)
Nuclearity (%)
Relation (%)
ENG
SPA
84.06
67.43
56.22
ENG
BSQ
86.22
68.24
53.28
SPA
BSQ
88.61
71.02
54.94
Surprisingly, results for the quantitative evaluation are slightly better when
Basque is involved in the comparison, which was not the case for the segmentation
Span agreement results (Table 14). Agreement, however, is higher for the
Nuclearity criterion when Basque is included (also the case for Span agreement
results shown earlier). Finally, the Relation agreement drops when Basque is
involved. We point out the source of this change and we discuss the results of the
Relation comparison in Sect. 2.4.2, where we present the final results of both
evaluation methods (Table 21).
3.2.2 Results of qualitative evaluation method
Table 16 and Table 17 include the final results for the entire corpus, which account
for agreement and disagreement in a qualitative way. In Table 16 results from the
123
A qualitative comparison method for rhetorical structures
agreement level obtained on the four types of measurements increases as the
relaxation of the agreement increases too, being RCA the most demanding
agreement, and R the more relaxed one.
In Table 18 we show summarized results of the three sources: total agreement
between annotators (Agreement), discrepancies because of annotation decisions
(Annotation Discrepancies) and discrepancies because of linguistic differences
(Translation Strategies).
As we observe in Table 18, the disagreement is higher when data of both A1
(English) and A2 (Spanish) are compared with A3 (Basque). That could be, as we
have interpreted from the results of Table 14, because English and Spanish are
typologically closer to each other than Basque is to either English or Spanish (Cenoz
Table 16 Qualitative evaluation results (F-measure): analysis of the sources of agreement
Classification
ENG-SPA
%
Agreement
ENG-BSQ
Gain (%)
%
SPA-BSQ
Gain (%)
40.33
%
Gain (%)
RCA
44.67
RC
49.34
4.67
42.66
2.33
42.33
45.66
3.33
RA
51.67
7
48.66
8.33
50.66
8.33
R
59.67
3.33
54.66
3.67
56.99
3
Table 17 Qualitative evaluation results (F-measure): analysis of the sources of disagreement
Classification
ENG-SPA (%) ENG-BSQ (%) SPA-BSQ (%)
Annotator-based discrepancies Nuclearity
N/N versus N/S
4.00
4.00
3.33
5.33
8.00
6.00
Attachment span
2.00
1.33
0.67
Relation
6.67
4.00
2.67
Similar relation
1.67
4.33
6.67
Mismatched relation 6.00
4.67
5.67
Specificity
0.67
4.33
5.33
No Match
6.33
6.67
4.67
4.67
3.33
4.67
Clause structure
1.67
1.67
1.33
Unit shift
1.33
2.67
1.67
Language-based discrepancies Marker change
Table 18 Qualitative evaluation results (F-measure): summary of results
Classification
ENG-SPA (%)
ENG-BSQ (%)
SPA-BSQ (%)
Agreement
59.67
54.66
56.99
Annotator-based discrepancies
32.67
37.33
35.01
Language-based discrepancies
7.67
7.67
7.67
123
M. Iruskieta et al.
2003). But this dispersion is not so large if we take into account the fact that there
are more Similar Relations and Specificity when A3’s data is compared with A1’s
and A2’s.
After aligning the contingency tables of the qualitative evaluation from all the
RS-structure in English, Spanish and Basque, we measured the agreement of
rhetorical relations with Fleiss Kappa (see Table 19) for assessing the reliability of
agreement between more than two annotators. The agreement attained across the
three annotators was moderate with a Kappa (Fleiss 1971) score of 0.484 (300
rhetorical relations, 15 texts). We show in Table 19 the agreement relation by
relation between the three annotators.
As we observe in Table 19, Fleiss’ Kappa measures show different degrees of
understanding rhetorical relations.
1.
2.
Almost perfect: PREPARATION.
Substantial: SUMMARY and CONCESSION.
Table 19 Qualitative evaluation results (Fleiss’ Kappa) for rhetorical relations
Relation
Kappa
z
p value
Preparation
0.851
25.528
0.000
Summary
0.712
21.361
0.000
Concession
0.705
21.155
0.000
List
0.554
16.629
0.000
Elaboration
0.531
15.933
0.000
Condition
0.525
15.763
0.000
Sequence
0.499
14.966
0.000
Restatement
0.424
12.723
0.000
Background
0.420
12.589
0.000
Circumstance
0.420
12.586
0.000
Contrast
0.376
11.272
0.000
Cause
0.352
10.552
0.000
Purpose
0.335
10.057
0.000
Result
0.301
9.017
0.000
Means
0.221
6.617
0.000
Conjunction
0.172
5.151
0.000
Motivation
0.136
4.084
0.000
Interpretation
0.080
2.390
0.017
Solutionhood
-0.011
-0.337
0.736
Justify
-0.009
-0.269
0.788
Antithesis
-0.008
-0.235
0.814
Evidence
-0.008
-0.235
0.814
Evaluation
-0.003
-0.100
0.920
Disjunction
-0.001
-0.033
0.973
Unless
-0.001
-0.033
0.973
123
A qualitative comparison method for rhetorical structures
3.
Moderate agreement: LIST, ELABORATION, CONDITION, SEQUENCE, RESTATEMENT,
and CIRCUMSTANCE.
Fair agreement: CONTRAST, CAUSE, PURPOSE, RESULT and MEANS.
Slight agreement: CONJUNCTION, MOTIVATION and INTERPRETATION.
No observed agreement for: ANTITHESIS, DISJUNCTION, EVALUATION, EVIDENCE,
19
JUSTIFY, SOLUTIONHOOD and UNLESS.
BACKGROUND
4.
5.
6.
Translation Strategies. In carrying out the comparison of rhetorical structures, we
observed some language differences. Some of them were produced when authors
translated from one language into another (translation strategy),20 and others were
the result of comparing rhetorical structure in a pairwise manner, for instance in
comparing English and Spanish with each other, when they are both translations of a
Basque source. The latter cannot be regarded as translation strategies, so we will
include only the first types under the umbrella term ‘translation shift’. And the
second type under the umbrella ‘different language forms’.
On the one hand, we do not analyze translation strategies which do not lead the
annotator to choose a different relation, as in Example (3); where in Basque the
rhetorical relation was made explicit with the marker (izan ere, ‘in fact’), but
remains the same relation, a CAUSE relation is in the A1 analysis.21
(3)
(a)
[In the recent past, a trend has been noted, and reported by many
researchers in the area of Serbian scientific terminology, of importing
borrowings of lexical and larger structural units from English into
specific scientific registers, rather that to opt for translations, calques,
etc.]3N [This corresponds closely to the fact that a consensus has been
reached among Serbian scientists of various orientations regarding the
status of English as the only language of scientific communication in the
last several decades.]4SCAUSE
(b) [Aurreko hamarkadetan, serbierako zientzia-arloko ikertzaile askok joera
bat nabaritu dute eta horren berri eman dute: ingeleseko unitate lexikalen
maileguak eta unitate-egitura luzeagoen maileguak hartzen dira zientziaerregistro zehatz baterako, itzulpenak edo kalkoak egin ordez.]3N [Izan
ere, iritzi ezberdinetako zientzialari serbiarrek adostasuna lortu dute eta
aurreko hamarkadetan ingelesari eman diote zientzia-komunikaziorako
hizkuntza bakarraren estatusa.]4SCAUSE TERM18_ENG
19
‘‘Values of agreement between A_e/1A_e (no observed agreement) and 1 (observed agreement =
1), with the value 0 signifying chance agreement (observed agreement = expected agreement).’’ (Artstein
and Poesio 2008, p. 559).
20
Catford (1965, pg. 73) defines translation shifts as ‘‘departures from formal correspondence in the
process of going from the SL to the TL’’ (from the Source Language to the Target Language).
Chesterman (1997) states that changes from original to translated text are due to a translation strategy.
21
Note that here there is another translation strategy (CSC hierarchical upgrading in Basque with a
coordination of two finite verbs lortu dute ‘½they achieve ½it’ and eman diote ‘½they give ½him’), which
is not under consideration due to harmonization process.
123
M. Iruskieta et al.
On the other hand, we do analyze all the directions (ENG [ SPA, ENG [ BSQ and
so on) in Table 20 and three types of translation differences that influence rhetorical
relations and reveal local translation strategies:
1.
2.
3.
1.
Relation signaling has a different configuration (Marker Change). Within
Marker Change, we found three subtypes:
(a) inclusion of a marker,
(b) exclusion of a marker, and
(c) changing a marker.
Differences because of the use of a distinct language configuration (Clause
Structure Change):
(a) hierarchical downgrading, and
(b) hierarchical upgrading.
Punctuation is used differently (Unit Shift):
(a) an independent sentence is integrated in another sentence, and
(b) a clause is translated in an independent sentence. We detail some of them
below.
Marker Change. In Example (4) a discourse maker (de ahí, ‘hence’) was not
translated from Spanish into either English or Basque. In English the marker
por ejemplo (‘for example’) was also elided and the punctuation changed (from
semicolon into colon). This is why annotators in English and Basque labelled
the relation ELABORATION; whereas in Spanish, the marker de ahí (‘hence’)
resulted in an annotation with the evidence label.
(4)
(a)
[Es más, desde cualquier lugar los términos son recopilados, comentados
y ponderados;]9N [de ahı́, por ejemplo, los apartados que encontramos en
muchos Webs en que se difunden glosarios de términos sobre Internet o
en que se exponen propuestas denominativas que los usuarios pueden
incluso votar.]10SEVIDENCE
(b) [Furthermore, terms can be compiled, discussed and assessed anywhere:]9N [many Web sites can be found which give glossaries of
Internet terms or propose names and even invite users to vote on
them.]10SELABORATION
(c) [Are gehiago, edozein tokitatik biltzen dira terminoak, baita komentatu
eta haztatu ere;]9N [adibidez, Interneti buruzko terminoen glosarioak
zabaltzen dira Web askotan, eta izendegietarako proposamenak egin ere
bai, eta erabiltzaileek botoa eman ahal izaten diete.]10SELABORATION
TERM38_SPA
2.
Clause Structure Change. In Example (5) the clauses under the relative used
in the original Spanish text were avoided in the same way in English and in
123
A qualitative comparison method for rhetorical structures
Basque (que si bien la ha enriquecido, al mismo tiempo ha puesto en cuestión
algunos de sus conceptos básicos, ‘that, although ½it has enriched it, ½it has also
called into question some of its basic concepts’), in favour of an adversative
coordination using a finite verb in English (but), and a conjunction coordination
(eta, ‘and’) and a finite verb in Basque (jarri ditu, ‘½it places ½them’). That was
the reason for A1 to annotate a CONTRAST relation, whereas A3 annotated a LIST
relation. The relative form22 analyzed here is a product of the harmonization
and it was annotated by A2 as an ELABORATION relation.
(5)
(a.)
[Todos estos factores, además de provocar un aumento cuantitativo de la
terminologı́a especializada, han implicado una ampliación de la
perspectiva del trabajo en terminologı́a,g6N fque si bien la ha
enriquecido, al mismo tiempo ha puesto en cuestión algunos de sus
conceptos básicos ð. . .Þ]711SELABORATION 23
(b.) [All these factors lead to an increase in the number of specialist terms
which enrich terminology]6NCONTRAST [but also call into question some
of its basic concepts ð. . .Þ]7NCONTRAST
(c.) [Alderdi horiek guztiek, espezialitateko terminologiaren gehikuntza
kuantitatiboa eragiteaz gain, terminologia lanen ikuspegia ere zabaldu
egin dute;]6NLIST [eta, egia bada ere ikuspegi berri horrek terminologia
aberastu egin duela esatea, zalantzan jarri ditu terminologiaren oinarrizko zenbait kontzeptu ð. . .Þ]7NLIST TERM19_SPA
3.
Unit Shift. A different punctuation can lead the annotator to interpret a
different relation. In the original text in Spanish in Example (6), the spans were
linked with comma, whereas in the English text the punctuation was changed,
using a period. The punctuation led A1 to consider a hypotactic relation
between the first and the following two spans.
(6)
(a)
[En esta comunicación, a partir de la experiencia en trabajos de
normalización de terminologı́a catalana, se planteará la necesidad social
de la normalización terminológica,]N12LIST [se comentarán algunas de
las dificultades con que se enfrenta y se apuntarán ideas para su enfoque
dentro de la sociedad actual.]N1314LIST
(b) [This paper looks, on the basis of experience in the standardisation of
terminology in Catalan, at the social need for standardisation of
terminology.]N12 [Some of the difficulties faced will be discussed, and
22
Again, this goes against the principles of our segmentation.
23
Note here the human annotation error which does not follow the modular and incremental annotation
that Pardo (2005) proposes.
123
M. Iruskieta et al.
ideas will be given for approaching this field in present day society.]S1314ELABORATION TERM19_SPA
We present, in Table 20, the influence of translation strategies and different
language forms more in depth.
It is worth mentioning that when English is the SL there are not so many
translation strategies (10.14 %) as when other languages are SL (Spanish: 23.19 %
and Basque: 34.78 %). Another interesting aspect is that the Marker Change
translation strategy is the most prominent one (MC: 34.78 % versus CSC: 15.94 %
and US: 17.39 %), and changes in discourse markers have an influence on rhetorical
annotation.24 These results are merely describing tendencies, because the corpus is
not big enough (although is comparable to other corpora in the literature, such as
Scott et al. (1998)). The results are sensitive to segmentation granularity or
harmonization decisions and to text characteristics (genre and domain). However
what is relevant is that the method presented here can describe and quantify
translation strategies.
3.2.3 Comparing quantitative and qualitative methodologies
To determine whether the proposed method is consistent, we compare the
quantitative results of the relation factor from both methods in Table 21. In this
table, we present the final results from both evaluation methods, providing the Fmeasure of relation factor.
We can highlight two findings in this comparison:
1.
2.
The qualitative method finds slightly higher agreement than the quantitative
method. The difference goes from almost 2 to 4 % when we compare results in
a pairwise manner.
Both methods show the same relative agreement rate per language pair. The
pair with the highest agreement corresponds to English-Spanish, second comes
the pair Spanish-Basque, and finally the pair English-Basque shows the lowest
agreement.
In the rhetorical analysis, unlike those we have achieved in the harmonization
(changes made in languages to carry out the alignment of discourse units), we see no
significant difference (Translation Strategies in Table 20) between languages
typologically more distant. It is worth noting, however, that for the closest
languages, the English-Spanish pair, the agreement in relation is higher. Languages
with more contact like the Spanish-Basque pair obtain better agreement than the
English-Basque pair (Table 21).
We see clear advantages to the use of the qualitative evaluation method. First of
all, with a qualitative evaluation, we measure inter-annotator agreement using only
RST relations. Relations and nuclearity are phenomena of a different nature, and we
believe they ought not to be included in the same factor. Secondly, the qualitative
evaluation clearly distinguishes the most relevant sources of disagreement; because
24
This phenomenon (marker change is the first reason to mismatch relations) is repeated when we
compare translated texts (TL) among them (MC 20.29 %, CSC 4,35 % and US 7.25 %).
123
Different language forms
1.45
2.90
CSC
US
Total 68.12
1.45
MC
2.90
1.45
–
2.90
2.90
4.35
1.45
4.35
7.25
4.35
4.35
10.14
2.90
1.45
11.59
31.88
0.00
2.90
14.49
4.35
1.45
4.35
2.90
–
1.45
ENG [ SPA (%) ENG [ BSQ (%) SPA [ ENG (%) SPA [ BSQ (%) BSQ [ ENG (%) BSQ [ SPA (%) ENG-SPA (%) ENG-BSQ (%) SPA-BSQ
Translation strategies
Table 20 Translation strategies and different language pairs
A qualitative comparison method for rhetorical structures
123
M. Iruskieta et al.
Table 21 Comparison of relation factor in quantitative and qualitative evaluation methods (F-measure)
Quantitative evaluation (%)
Qualitative evaluation (%)
ENG-SPA
56.22
59.67
ENG-BSQ
53.28
54.66
SPA-BSQ
54.94
56.99
of that, results are more reliable. The translation of discourse structure from one
language to another does not result in a one-to-one mapping of relations. As Marcu
(2000a) has mentioned, sometimes a particular rhetorical structure has to be
translated as a different structure. Moreover, translation strategies can affect the
rhetorical structure and annotation, and the qualitative method presented here could
be used to identify and measure these translation strategies.
4 Conclusions and further work
The methodology we have proposed has two main implications for RST theory and
for annotation methodology. First of all, in terms of RST theory, we have shown
that it is possible to conduct cross-linguistic studies using the same set of principles.
In our study we have shown that, although RST structures may not be exactly the
same across languages, they do show a large similarity. Secondly, we have provided
a clear and detailed method to identify where structures differ. Thirdly, the
annotated files are available to anyone who wishes to use them and on our website25
the tagged multilingual corpus can be consulted, as for example: (1) the rhetorical
structure of a text (in RS3 format) and its image (in JPG format); (2) all instances of a
selected rhetorical relation in three languages; (3) discourse units of a text in each
language or aligned in three languages.
Ours is, to our knowledge, the first study that provides a rigorous qualitative
methodology for comparison of rhetorical structures, which solves the deficiencies
of quantitative evaluations and provides a qualitative description of agreement and
disagreement. This method distinguishes and locates translation strategies when
those strategies are the sources of annotator disagreement, as opposed to simple
annotator discrepancies. The methodology helps determine whether the same
passage in different languages has different RST structures because those structures
correspond to different applications of the theory, or whether the discrepancy in
RST structures is due to different linguistic realizations (due to translation
strategies, broadly understood).
The study has some limitations with regard to the source of the translation
differences that the analysis reveals. We believe that in order to detect these sources
a translation theory ‘‘must include both a descriptive and an evaluative element’’, as
Chesterman (1993) suggests, so that we can decide whether translation strategies
may or may not be well motivated. We have presented some suggestions for the
25
http://ixa2.si.ehu.es/rst.
123
A qualitative comparison method for rhetorical structures
translation differences that the analysis evidenced, showing that typological
differences between the languages affected mostly segmentation. More detail,
informed by a rigorous translation theory, is necessary, but is beyond the scope of
this paper.
Our results show that RST, in conjunction with our methodological proposal for
the comparison of RST annotations, are valid tools for the study of translated
corpora. The results of our corpus analysis provide some evidence that, in
segmentation, the linguistic distance calculated by change in the harmonization
process is very small between languages from the same family such as EnglishSpanish and it is large between languages from distinct families such as SpanishBasque and English-Basque. Surprisingly, the dispersion in relation agreement
caused by translation strategies was very small when comparing English-Basque
and Spanish-Basque with English-Spanish. In the same line, the linguistic distance
in rhetorical relations, calculated as the F-score result when comparing RST
annotations, is not as large as the segmentation differences. It appears that there is
more dispersion in segmentation than in rhetorical relations; this may be due to the
fact that there is more distance at the level of clause linking than at the level of
discourse relational structure. It is worth noting, however, that each language is
affected by a particular translation strategy in this corpus.
Although the results obtained by both methods in the annotations for different
languages show that there are different interpretations, this is not due to interlingual
differences. The problem of annotation subjectivity arises also when three
annotators analyze the same text in a language: this problem is even more
important when the annotators do not have the same training (although in our
experiment the three annotators started their annotation from the same departure
criteria). As we said, the purpose of this paper is to present a methodology to
compare RS-trees and not to describe the structure of text in the three languages. To
see a description of those texts and a detailed work in these three languages, we
recommended consulting the corpora developed by the authors in these three
languages (English SFU corpus26 (Taboada and Renkema 2008), Spanish RST
TreeBank27 (da Cunha et al. 2011b) and Basque RST TreeBank28 (Iruskieta et al.
2013a)). We are aware that in this work we do not account for the problem of
multiple relations in RST (Taboada and Mann 2006b; Marcu 2000b) or all the
possibilities comparing RS-trees in parallel corpora.
The qualitative evaluation is in certain respects more complex than Marcu’s
quantitative evaluation, which has been automated by Maziero and Pardo (2009).
Despite its complexity, it solves some inherent problems of the quantitative
evaluation and it has advantages when describing the sources of disagreement.
We plan to perform two tasks as future work. First of all, we will carry out a
larger RST multilingual corpus analysis, but limited to a smaller number of
rhetorical relations, with the objective of detecting translation strategies in order to
improve machine translation discourse tasks. Second, we will carry out an automatic
26
SFU corpus is available at http://www.sfu.ca/*mtaboada/download/downloadRST.html.
27
RST Spanish TreeBank is available at http://corpus.iingen.unam.mx/rst/corpus_en.html.
28
Basque RST TreeBank is available at http://ixa2.si.ehu.es/diskurtsoa/en/.
123
M. Iruskieta et al.
implementation of the qualitative rhetorical evaluation that we propose in our work,
which will be valid for monolingual (Iruskieta et al. 2013a) and multilingual
annotation, so that it can be used by all the scientific community working on RST.
Acknowledgments This work has been partially financed by the Spanish projects RICOTERM 4
(FFI2010-21365-C03-01) and APLE 2 (FFI2012-37260), and a Juan de la Cierva Grant (JCI-2011-09665)
to Iria da Cunha. Maite Taboada was supported by a Discovery Grant from the Natural Sciences and
Engineering Research Council of Canada (261104-2008). Mikel Iruskieta was supported by the following
projects: OPENMT-2 (TIN2009-14675-C03-01) [Spanish Ministry], Ber2Tek (IE12-333) [Basque
Government] and IXA group (GIU09/19) [University of the Basque Country]. We would like to thank
the anonymous reviewers for their comments and suggestions, Nynke van der Vliet for her feedback on
the evaluation method, Esther Miranda for designing the website, and Oier Lopez de Lacalle for helping
with the scripts to calculate the statistics.
Appendix: Discourse segmentation details
The first step in analyzing texts under RST consists of segmenting the text into
spans. Exactly what a span is, under RST, and more generally in discourse, is a welldebated topic. RST Mann and Thompson (1988) proposes that spans, the minimal
units of discourse—later called elementary discourse units (EDUs) (Marcu
2000a)—are clauses, but that other definitions of units are possible:
The first step in analyzing a text is dividing it into units. Unit size is arbitrary,
but the division of the text into units should be based on some theory-neutral
classification. That is, for interesting results, the units should have independent
functional integrity. In our analyzes, units are essentially clauses, except that
clausal subjects and complement and non-restrictive relative clauses are
considered as part of their host clause units rather than as separate units.
(Mann and Thompson 1988, p. 248)
This definition is the basis of our work. From our point of view, adjunct clauses
stand in clear rhetorical relations (cause, condition, concession, etc.). Complement
clauses, however, have a syntactic, but not discourse, relation to their host clause.
Complement clauses include, as Mann and Thompson (1988) point out, subject and
object clauses, and restrictive relative clauses, but also embedded report complements, which are, strictly speaking, also object clauses.
Other possibilities for segmentation exist; one of the better-known ones is the
proposal by Carlson et al. (2003) for segmentation of the RST Discourse Treebank
(Carlson et al. 2002). Carlson et al. (2003) propose a much more fine-grained
segmentation, where report complements, relative clauses and appositive elements
constitute their own EDUs.
In our work three annotators segmented the EDUs of each corpus (A1 segmented
English texts, A2 segmented Spanish texts, and A3 segmented Basque texts). These
annotators are experts on RST, since they have been researching in this field since
years ago, and they have participated in several projects related to the design and
elaboration of RST corpora in the three languages of this work. Annotators
performed this segmentation task separately and without contact among them. In
123
A qualitative comparison method for rhetorical structures
our segmentation, we follow then the general guidelines proposed by Mann and
Thompson (1988), which we have operationalized for this paper. We detail the
principles below.
Every EDU Should Have a Verb
In general, EDUs should contain a (finite) verb. The main exception to this rule is
the case of titles, which are always EDUs, whether they contain a verb or not.
Non-finite verbs form their own EDUs only when introducing an adjunct clause
(but not a modifier clause, as we will see below). In (7), the non-finite clause
Focussing on less widely... is an independent EDU, because it is an adjunct clause.
Note that in both Spanish and Basque the same proposition was translated as an
independent sentence.
(7)
(a)
[Focussing on less widely used and taught languages (LWUTLs)
including Irish,] [the VOCALL partners are compiling multilingual
glossaries of technical terms in the areas of computers, office skills and
electronics] [and this involves the creation of a large number of new Irish
terms in the above areas.]
(b) [El proyecto está enfocado hacia lenguas minoritarias en cuanto al uso y
enseñanza, incluido el irlandés.] [El proyecto VOCALL estáen proceso
de recopilación de un glosario plurilingüe de términos técnicos de las
áreas de informática, secretariado y construcción,] [y esto supone la
creación de una larga serie de nuevos términos en irlandés, en las áreas
mencionadas.]
(c) [Gutxi erabiltzen eta irakasten diren hizkuntzetan kontzentratzen da
proiektua (LWUTL), irlandera barne.] [Informatika, bulego-lana eta
eraikuntzako arloetako termino teknikoen glosario eleanizduna biltzen ari
da VOCALL,] [eta horrek esan nahi du arlo horietako irlanderazko
termino berri ugari sortzen ari dela.] TERM23_ENG
In some cases, a prepositional phrase (especially one containing a nominalized verb)
in one language was realized as an independent clause in another. The final decision
in such cases is typically to segment minimally, that is, to unify the segmentation
across the three languages, so that the language with the fewer segments determines
how the texts in the other languages have to be segmented. See also Sect. 3.1.1, on
harmonization of the segmentation, for more examples of our final decisions across
the three languages.
Coordination and Ellipsis. Coordinated clauses are separated into two segments,
including cases where the subject is elliptical in the second clause. In Spanish and
Basque, both pro-drop languages, this is in fact the default for both first and second
clause, and therefore we see no reason why a clause with a pro-drop subject cannot
be an independent unit. We follow the same principle for English. In (8), the first
two EDUs in Spanish are coordinated with an elliptical subject in both cases,
referring to the authors (venimos traduciendo, ‘½we have been translating’ and
queremos expresar, ‘½we wish to indicate’). They constitute separate EDUs. In the
English and Basque versions, the two clauses are expressed as separate sentences.
123
M. Iruskieta et al.
(8)
(a)
[To attain this goal we have been translating doctrinal texts in law at the
University of Deusto since 1994.] [We wish to indicate the difficulties we have
had over the years and also our achievements,] [if there can be said to be any.]
(b) [Para poder alcanzar ese objetivo en la Universidad de Deusto venimos
traduciendo textos doctrinales del campo del Derecho desde 1994] [y
queremos expresar las dificultades que hemos tenido a lo largo de estos
años y, ası́mismo, también los logros conseguidos,] [si es que realmente
los ha habido.]
(c) [Xede hori iristeko, 1994. urteaz geroztik, Deustuko Unibertsitatean
Zuzenbidearen inguruko testu doktrinalak itzultzen dihardugu.] [Esperientzia horretan izandako zailtasunak eta,] [halakorik izanez gero,]29
[lorpenak ere azaldu nahi ditugu.] TERM25_BSQ
Coordinated verb phrases (VPs) or verbs do not constitute their own EDUs. We
differentiate coordinated clauses from coordinated VPs because the former can be
independent clauses with the repetition of a subject; the latter, in the second part of
the coordination, typically contain elliptical verbal forms, most frequently a finite
verb or modal auxiliary.
Relative, Modifying and Appositive Clauses. We do not consider that relative
clauses (restrictive or non-restrictive), clauses modifying a noun or adjective, or
appositive clauses constitute their own EDUs. We include them as part of the same
segment together with the element that they are modifying. This departs from RST
practice, where (restrictive) relative clauses are often independent spans, as seen in
many of the examples in the original literature and the analyzes on the RST web site
(Mann and Thompson 1988; Mann and Taboada 2010). We found that relative
clauses and other modifiers often lead to truncated EDUs, resulting in repeated use
of the Same-unit relation (see Truncated EDUs in 5 section), and thus decided that it
was best to not elevate them to the status of independent segments.
An example is presented in (9), where the relative clause is in parentheses in the
Spanish original. Note, however, that the coordinated clauses (with an elliptical
subject in all cases) are independent segments, as explained above. In Basque, on
the other hand, the relative clause is translated as an independent clause with a finite
verb (mugatzen da, ‘[it] is limited to’). We have not segmented it in Basque, to
agree with the other two languages.
(9)
½. . . [Internet terminology extends beyond the bounds of its specialist
field (which by definition is part of the lexicon of science and
technology)] [and breaks into general language.]
(b) ½. . . [la terminologı́a de Internet traspasa los lı́mites del área de
especialidad (a la que se circunscribe por definición el léxico cientı́fico y
técnico)] [e irrumpe en la lengua de uso general,] ½. . .
(c) ½. . . [espezialitateko eremuaren mugak gainditzen dituela Interneteko
terminologiak (espezialitatera mugatzen da, definizioz, lexiko zientifiko
(a)
29
Truncated EDU. English translation: ‘if there can be said to be any’ (see Sect. 5).
123
A qualitative comparison method for rhetorical structures
eta teknikoa),] [eta erabilera orokorreko hizkeran sartzen dela indartsu;]
½. . . TERM38_SPA
Parentheticals. The same principle applies to parentheticals and other units
typographically marked as separate from the main text (with parentheses or dashes).
They do not form an individual span if they modify a noun or adjective as in
Example 10, but they do if they are independent units, with a finite verb. Such is the
case in (11), with a full sentence in the parenthetical unit (in English, composed of
three finite clauses: can... be represented, is and are).
(10)
(a)
The analysis of the data at hand—international terms most of which
have not yet been standardized in Serbian—indicate that a hierarchy of
criteria for evaluating the terms, (...). TERM18_ENG
(11)
(a)
[The design and management of terminological databases pose
theoretical and methodological problems] [(how can a term be
represented?] [Is there a minimum representation?] [How are terms to
be classified?),] ð. . .Þ
(b) [Efectivamente, el diseño y la gestión de las bases de datos terminológicos plantean problemas diversos tanto de ı́ndole teórica y
metodológica] [(>cómo se representa un término?,] [>existe una
representación mı́nima?,] [>cómo se clasifican los términos?)] ð. . .Þ
(c) [Hala da, terminologiako datu-baseak diseinatzeak eta kudeatzeak
hainbat arazo dakar bai teoria eta metodologiaren aldetik] [(nola
adierazi terminoa?] [Ba al da gutxieneko adierazpenik?] [Nola sailkatu
terminoak?),] ð. . .Þ TERM29_SPA
Reported Speech. We believe that reported and quoted speech do not stand in
rhetorical relations to the reporting units that introduce them, and thus should not
constitute separate EDUs, also following clear arguments presented elsewhere (da
Cunha and Iruskieta 2010; Stede 2008a). This is in contrast to the approach in the
RST Discourse Treebank (Carlson et al. 2003), where reported speech (there named
ATTRIBUTION) is a separated EDU. There are, in any case, no examples of reported
speech in our corpus.
Truncated EDUs. In some cases, a unit contains a parenthetical or inserted unit,
breaking it into two separate parts, which do not have any particular rhetorical
relation between each other. In those cases, we make use of a non-relation label,
Same-unit, proposed for the RST Discourse Treebank (Carlson et al. 2003).
We see one such example in (11) above. The element that corresponds to the
third unit in English is, in fact, inserted in the middle of the second unit in Basque.
In order to align or harmonize segmentation and to preserve the integrity of that
unit, we use the Same-unit (non) relation, as shown in Fig. 8, which follows the
Basque word order.
123
M. Iruskieta et al.
Fig. 8 Example of a Same-unit (non) relation
Once our segmentation criteria were established and the three annotators carried
out the segmentation, the three segmentations were compared in terms of precision
and recall. In this way, we quantified agreement and disagreement across
segmentations. Moreover, we analyzed the main causes of the disagreements.
Results are shown in Sect. 3. After the segmentation agreement evaluation, we
harmonized the segmentation, ensuring that units were comparable across the
languages. At this point, we also calculated linguistic distance between the pairs of
languages, We understand linguistic distance as ‘‘the extent to which languages
differ from each other’’ (Chiswick and Miller 2005, pg. 1). Although this concept is
well known among linguists, there is not a single measure to evaluate this distance
Chiswick and Miller (2005). In our work, in order to measure this distance we
calculated which language required the most changes in the harmonization process.
This harmonization process was necessary to start out the analysis with similar
units, and to avoid confusing analysis disagreement and segmentation agreement.
Marcu et al. (2000) and Ghorbel et al. (2001) also align (which we termed
harmonize) their texts, decreasing the granularity of their segmentation to avoid
complexity. With this decision, we lose some rhetorical information at the most
detailed level of the tree. This does not, however, affect higher levels of tree
structure. The results of this harmonization are shown in Sect. 3.1.
References
Abelen, E., Redeker, G., & Thompson, S. A. (1993). The rhetorical structure of US-American and Dutch
fund-raising letters. Text, 13(3), 323–350.
Artstein, R., & Poesio, M. (2008). Inter-coder agreement for computational linguistics. Computational
Linguistics, 34(4), 555–596.
Baker, M. (2004). A corpus-based view of similarity and difference in translation. International Journal
of Corpus Linguistics, 9(2), 167–193.
Bateman, J. A., & Rondhuis, K. J. (1997). Coherence relations: Towards a general specification.
Discourse Processes, 24(1), 3–49.
123
A qualitative comparison method for rhetorical structures
Carlson, L., Okurowski, M. E., & Marcu, D. (2002). RST Discourse Treebank, LDC2002T07 [Corpus].
Philadelphia, PA: Linguistic Data Consortium.
Carlson, L., Marcu, D., & Okurowski, M. E. (2003). Building a discourse-tagged corpus in the framework
of rhetorical structure theory. In van Kuppevelt, C. J. Jan & R. W. Smith (Eds.), Current and new
directions in discourse and dialogue (pp. 85–112). Berlin: Springer.
Catford, J. C. (1965). A linguistic theory of translation: An essay in applied linguistics (Vol. 8). New
York: Oxford University Press.
Cenoz, J. (2003). The role of typology in the organization of the multilingual lexicon. In J. Cenoz, B.
Hufeisen & U. Jessner (Eds.), The multilingual lexicon (pp. 103–116), New York: Springer.
Chesterman, A. (1993). From ‘is’ to ‘ought’: Laws, norms and strategies in translation studies. Target,
5(1), 1–20.
Chesterman, A. (1997). Memes of translation: The spread of ideas in translation theory (Vol. 22).
Amsterdam and Philadelphia: Benjamins.
Chiswick, B. R., & Miller, P. W. (2005). Linguistic distance: A quantitative measure of the distance
between english and other languages. Journal of Multilingual and Multicultural Development,
26(1), 1–11.
Cristea, D., Ide, N., & Romary, L. (1998). Veins theory: A model of global discourse cohesion and
coherence. In C. Boitet & P. Whitelock (Eds.), 17th international conference on Computational
linguistics (Vol. 1 pp. 281–285). Montreal, Canada: Association for Computational Linguistics.
Cui, S. (1986). A comparison of English and Chinese expository rhetorical structures. Ph.D. thesis,
UCLA.
da Cunha, I., & Iruskieta, M. (2010). Comparing rhetorical structures in different languages: The
influence of translation strategies. Discourse Studies, 12(5), 563–598.
da Cunha, I., Torres-Moreno, J. M., & Sierra, G. (2011a). On the Development of the RST Spanish
Treebank. In 5th Linguistic annotation workshop. 49th annual meeting of the association for
computational linguistics, ACL (pp. 1–10). Portland, Oregon, USA.
da Cunha, I., Torres-Moreno, J. M., Sierra, G., Cabrera-Diego, L. A., Castro-Rolón, B. G., & RollandBartilotti, J. M. (2011b). The RST Spanish Treebank On-line Interface. In International conference
recent advances in NLP (pp. 698–703), Bulgaria.
Delin, J., Hartley, A. F., Paris, C., Scott, D. R., & Linden, K. V. (1994). Expressing procedural
relationships in multilingual instructions. In Seventh International Workshop on Natural Language
Generation (pp. 61–70), Association for Computational Linguistics.
Delin, J., Hartley, A. F., & Scott, D. R. (1996). Towards a contrastive pragmatics: Syntactic choice in
English and French instructions. Language Sciences, 18(3–4), 897–931.
Egg, M., & Redeker, G. (2010). How complex is discourse structure? In Proceedings of the 7th
international conference on language resources and evaluation (LREC 2010) (pp. 1619–1623),
Valletta, Malta.
Fetzer, A., & Johansson, M. (2010). Cognitive verbs in context. A contrastive analysis of English and
French argumentative discourse. International Journal of Corpus Linguistics, 15(2), 240–266.
Fleiss, J. L. (1971). Measuring nominal scale agreement among many raters. Psychological Bulletin,
76(5), 378–382.
Flowerdew, J. (2010). Use of signalling nouns across l1 and l2 writer corpora. International Journal of
Corpus Linguistics, 15(1), 36–55.
Fung, P. (1995). Compiling bilingual lexicon entries from a non-parallel English–Chinese corpus. In 3rd
workshop on very large Corpora, (Vol. 78, pp. 173–183). Boston, MA.
Ghorbel, H., Ballim, A., & Coray, G. (2001). ROSETTA: Rhetorical and semantic environment for text
alignment. In: Corpus Linguistics, Lancaster University (UK) (pp. 224–233).
Gomez, X., & Simoes, A. (2009). Parallel corpus-based bilingual terminology extraction. In 8th
international conference on terminology and artificial intelligence Toulouse.
Granger, S. (2003). The corpus approach: A common way forward for Contrastive Linguistics and
Translation Studies (pp. 17–29). Rodopi, Corpus-based approaches to contrastive linguistics and
translation studies. Amsterdam/New York.
House, J. (2004). Explicitness in discourse across languages. Neue Perspektiven in der Übersetzungs-und
Dolmetschwissenschaft (pp. 185–208), Bochum: AKS.
Iruskieta, M., Aranzabe, M. J., Dı́az de Ilarraza, A., Gonzalez, I., Lersundi, M., & Lopez de la Calle, O.
(2013a). The RST Basque TreeBank: An online search interface to check rhetorical relations. In 4th
workshop RST and discourse studies, Brasil.
123
M. Iruskieta et al.
Iruskieta, M., Dı́az de Ilarraza, A., & Lersundi, M. (2013b). Establishing criteria for RST-based discourse
segmentation and annotation for texts in Basque. Corpus Linguistics and Linguistic Theory, 1–32.
Kanté, I. (2010). Mood and modality in finite noun complement clauses: A French-English contrastive
study. International Journal of Corpus Linguistics, 15(2), 267–290.
Koehn, P. (2005). Europarl: A parallel corpus for statistical machine translation. In: MT summit, Phuket,
Thailand.
Kong, K. C. C. (1998). Are simple business request letters really simple? A comparison of Chinese and
English business request letters. Text & Talk, 18(1), 103–141.
Mann, W. C., & Taboada, M. (2010). RST web-site. http://www.sfu.ca/rst/. Accessed 30 September 2012.
Mann, W. C., & Thompson, S. A. (1988). Rhetorical structure theory: Toward a functional theory of text
organization. Text-Interdisciplinary Journal for the Study of Discourse, 8(3), 243–281.
Marcu, D. (2000a). The rhetorical parsing of unrestricted texts: A surface-based approach. Computational
Linguistics, 26(3), 395–448.
Marcu, D. (2000b). The theory and practice of discourse parsing and summarization. Cambridge: MIT
press.
Marcu, D., Carlson, L., & Watanabe, M. (2000). The automatic translation of discourse structures. In 1st
North American chapter of the Association for Computational Linguistics conference (pp. 9–17),
Seattle (USA): Morgan Kaufmann Publishers.
Maxwell, M. (2010). Limitations of corpora. International Journal of Corpus Linguistics, 15(3), 379–383.
Maziero, E. G., & Pardo, T. A. S. (2009). Automatização de um método de avaliação de estruturas
retóricas. In: RST Brazilian meeting, São Paulo, Brazil.
Mitocariu, E., Anechitei, D. A., & Cristea, D. (2013). Comparing discourse tree structures (pp. 513–522).
Berlin: Springer. Computational Linguistics and Intelligent Text Processing.
Mohamed, A. H., & Omer, M. R. (1999). Syntax as a marker of rhetorical organization in written texts:
Arabic and English. International Review of Applied Linguistics in Language Teaching (IRAL),
37(4), 291–305.
Morin, E., Daille, B., Takeuchi, K., & Kageura, K. (2007). Bilingual terminology mining-using brain, not
brawn comparable corpora. In Annual meetings ACL (Vol. 45, pp. 664–671). Prague.
Mortier, L., & Degand, L. (2009). Adversative discourse markers in contrast: The need for a combined
corpus approach. International Journal of Corpus Linguistics, 14(3), 338–366.
O’Donnell, M. (2000). RSTTool 2.4: A markup tool for rhetorical structure Theory. In First international
conference on natural language generation INLG’00 (Vol. 14, pp. 253–256). Mitzpe Ramon: ACL.
Pardo, T. A. S. (2005). Métodos para análise discursiva automática. Ph.D. thesis, Instituto de Ciências
Matemáticas e de Computação, São Carlos-SP: Universidade de São Paulo.
Ramsay, G. (2000). Linearity in rhetorical organisation: A comparative cross-cultural analysis of
newstext from the People’s Republic of China and Australia. International Journal of Applied
Linguistics, 10(2), 241–258.
Ramsay, G. (2001). Rhetorical styles and newstexts: A contrastive analysis of rhetorical relations in
Chinese and Australian news-journal text. ASAA E-Journal of Asian Linguistics and Languageteaching, 1(1), 1–22.
Salkie, R., & Oates, S. L. (1999). Contrast and concession in French and English. Languages in Contrast,
2(1), 27–56.
Sarjala, M. (1994). Signalling of reason and cause relations in academic discourse. Anglicana Turkuensia,
13, 89–98.
Scott, D. R., Delin, J., & Hartley, A. F. (1998). Identifying congruent pragmatic relations in procedural
texts. Languages in Contrast, 1(1), 45–82.
Soricut, R., & Marcu, D. (2003). Sentence level discourse parsing using syntactic and lexical information.
In 2003 conference of the North American Chapter of the Association for Computational Linguistics
on human language technology (Vol. 1, pp. 149–156). Association for Computational Linguistics.
Stede, M. (2008a). Disambiguating rhetorical structure. Research on Language and Computation, 6(3),
311–332.
Stede, M. (2008b). RST revisited: Disentangling nuclearity (pp. 33–57). Amsterdam and Philadelphia:
John Benjamins. ‘Subordination’ versus ‘coordination’ in sentence and text.
Taboada, M. (2004a). Building coherence and cohesion: Task-oriented dialogue in English and Spanish.
Amsterdam and Philadelphia: John Benjamins.
Taboada, M. (2004b). Rhetorical relations in dialogue: A contrastive study (pp. 75–97), Amsterdam and
Philadelphia: John Benjamins. Discourse across Languages and Cultures.
123
A qualitative comparison method for rhetorical structures
Taboada, M., & Mann, W. C. (2006a). Applications of rhetorical structure theory. Discourse Studies,
8(4), 567–588.
Taboada, M., & Mann, W. C. (2006b). Rhetorical structure theory: Looking back and moving ahead.
Discourse Studies, 8(3), 423–459.
Taboada, M., & Renkema, J. (2008). Discourse relations reference corpus. Simon Fraser University and
Tilburg University. http://www.sfu.ca/rst/06tools/discourse_relations_corpus.html. Accessed 30
September 2012
Trask, R. L. (1997). The history of Basque. London: Routledge.
Usoniene, A., & Soliene, A. (2010). Choice of strategies in realizations of epistemic possibility in English
and Lithuanian: A corpus-based study. International Journal of Corpus Linguistics, 15(2), 291–316.
UZEI and HAEE-IVAP. (1997). International congress on terminology. Donostia and Gasteiz: UZEI;
HAEE-IVAP.
van der Vliet, N. (2010). Inter annotator agreement in discourse analysis. http://www.let.rug.nl/
*nerbonne/teach/rema-stats-meth-seminar/.
Wu, D., & Xia, X. (1994). Learning an English–Chinese lexicon from a parallel corpus. In First
conference of the AMTA (pp. 206–213). Citeseer, Columbia.
Xiao, R. (2010). How different is translated Chinese from native Chinese? A corpus-based study of
translation universals. International Journal of Corpus Linguistics, 15(1), 5–35.
123
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

Related manuals

Download PDF

advertisement