A qualitative comparison method for rhetorical structures: identifying different discourse structures in multilingual corpora

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.

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 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:

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 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, ‘$[$we$]$ have been translating’ and queremos expresar, ‘$[$we$]$ wish to indicate’). They constitute separate EDUs. In the English and Basque versions, the two clauses are expressed as separate sentences.

• (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, asímismo, 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)

  • (a) $[\dots ]$ [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) $[\dots ]$ [la terminología de Internet traspasa los límites del área de especialidad (a la que se circunscribe por definición el léxico científico y técnico)] [e irrumpe en la lengua de uso general,] $[\dots ]$

  • (c) $[\dots ]$ [espezialitateko eremuaren mugak gainditzen dituela Interneteko terminologiak (espezialitatera mugatzen da, definizioz, lexiko zientifiko eta teknikoa),] [eta erabilera orokorreko hizkeran sartzen dela indartsu;] $[\dots ]$ 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?),] $(\ldots )$

  • (b) [Efectivamente, el diseño y la gestión de las bases de datos terminológicos plantean problemas diversos tanto de índole teórica y metodológica] [(¿cómo se representa un término?,] [¿existe una representación mínima?,] [¿cómo se clasifican los términos?)] $(\ldots )$

  • (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?),] $(\ldots )$ 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.

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.