Back to Jennifer’s homepage | AI RUG Homepage |

The Contribution of Contrast in Context

NWO funded VENI project

Principle Investigator: Jennifer Spenader

Contrast as a key organizing principle of discourse

Since the Pythagoreans identified the ten fundamental oppositions (Ogden 1932), contrast and contrasting has played a leading role in philosophy, and seems to be a basic human cognitive skill. Therefore it is not strange that some category of contrast is included in almost all major taxonomies of rhetorical relations (e.g. Hobbs 1990, Martin 1992, Mann & Thompson, 1988 and Asher & Lascarides, 2003), evidence that many consider contrast to be a major organizing principle of information in discourse.

Examples of clauses joined with explicit contrastive markers like but are the clearest cases:

(1) [It's raining]1 but [I'm taking an umbrella]2.

The first conjunct is generally analyzed as implying something that an implication of the second conjunct will contradict, here something like “I will get wet” vs. “I won't get wet”. However, even for this well-studied type of contrast, many issues remain unexplored. Greater computational power means rhetorical relations and information structure can be taken into account in applications such as question answering, information retrieval, and in producing natural synthetic speech. But our current state of knowledge about contrast doesn’t support the integration of the information contributed by contrastive marking in an analysis in any practical way because it leaves too many basic questions unanswered.

There is a significant need for basic information about the purpose of contrastive relations and how they rhetorically relate with the rest of a dialogue. Additionally, corpus study shows that most cases of contrast are not even explicitly marked (Carlson et al. 2001):

(2) China violates human rights in Tibet daily and the United States continues to trade with them.

Automatic identification of unmarked contrast requires first developing an explicit description of marked contrast. Furthermore, in order to realistically deal with contrast in multi-speaker dialogue, such as would be necessary in a question-answering system, we also need a formal model of contrast that takes both speaker and hearer perspectives into account.

Three main research questions address these issues.

Q1 What is the contribution of contrastive relations in discourse?

Earlier work has made some tentative claims relating to the contribution of contrast. Winter & Rimon (1994) argue that the point of contrastive marking is to contribute the information implied by the second conjunct, and Spooren (1989) has shown

experimentally that indeed, the second conjunct is often interpreted as the speaker's opinion. Thus in (3), subjects generally believe that Jill intends to choose Lars:

(3) Jack: Will you choose Sven or Lars for your basketball team?

Jill: Sven's fast but Lars is taller.

In short, the second conjunct makes the key contribution. But what is the rhetorical contribution of the second conjunct in the context? And if the second conjunct is most important, what is the purpose of the first conjunct?

The two main research traditions, (A) rhetorical structural research and (B) information structural research, approach these questions very differently.

Rhetorical structural research (A) such as Asher and Lascarides (2003) with Structured Discourse Representation Theory (SDRT), Rhetorical Structure Theory (RST) (Mann and Thompson 1988), or the work of Kehler (2000) have viewed contrast as a coherence relation between two discourse segments, on a par with relationships such as Explanation (in (4)) and Narration (in (5)):

(4) John isn't coming. He's sick.

(5) John entered the room. He sat down.

A Contrast relation is then related to the rest of the discourse via some other rhetorical relation:

(6) The butler must be the murderer. He doesn't have a motive, but his fingerprints were on the gun.

In (6) the second sentence contributes a contrastive relation that functions as a type of Evidence to the statement made in the first sentence.

Preliminary corpus work (Spenader, 2005) suggests that contrast relations frequently serve as arguments to Evidence relations, as well as often being arguments to Narration. Further, contrast relations often served to deny, or correct something said by a previous speaker:

(7) John: Lars is a linguistics professor.

Jill: Lars does work at the university, but he’s a post-doc.

This function is predicted by Spenader (2004a) and Maier & Spenader (2004) which suggested that contrast can be analyzed on a semantic level in the same way as the speech-act of a denial, the only difference being that contrast involves world-knowledge based inferences. This analysis effectively argues that a speaker may use contrast to manage what each speaker knows and/or agrees on in a conversation (i.e. align the common ground). Additionally, this analysis partly explains the function of the first conjunct. In (7), Jill’s first conjunct serves two purposes: 1) it delimits correct information from incorrect, and 2) it softens Jill’s correction of John by conceding that some information was correct.

Information structural research (B) hasn't given a clear account of contrast’s function in discourse either. Krifka (1999) and Umbach (2001) discuss contrastive conjuncts as each contributing a contrastive topic to the discourse which are then interpreted as partial answers to explicit or implicit questions in the previous context. Whether a particular type of question consistently occurs with contrast relations has yet to be discussed.

My hypothesis is that there is a strong tendency for Contrast in dialogue to be consistently used as Evidence, for a speaker to more strongly support their previous point. Cases where contrast resembles denial might be reducible to a limiting type of evidence; i.e. just making your point rather than giving information that indirectly supports it serves the same function. If corpus study validates this, the information can be used to automatically identify unmarked contrast relations and Evidence relations, as well as contributing significantly to a theory of a function of contrast.

Of the work cited above only Spooren (1989), Spenader (2004a) and Spenader (2005) are empirically informed. I will use corpus data to find initial answers to Q1 looking primarily at explicitly marked cases of contrast. Then, by experiments I will study (i) Do subjects perceive contrastive topics more clearly if the relation to the context is Narration rather than Evidence?, (ii) Do speakers arrive at similar hidden inferences when asked to interpret contrast relations with, and (iii) without a context?

Q2 How does the top-down view of contrast from rhetorical structure research (A) relate to the bottom-up view of contrast from information structural research (B)?

Both (A) and (B) are context-dependent theories used to analyze coherence creating devices in discourse, but it is not obvious how these two theoretical perspectives relate or when it is appropriate to utilize each theory (see Webber et al 2003 and Stede 2004 for a recent discussion). This uncertainty is particularly apparent in the treatment of contrast, where even in purely rhetorical structural work (i.e. in SDRT’s treatment, Asher & Lascarides 2003, Asher 2004) references are made to the information structural concepts of themes and topics, concepts never invoked to characterize most other key rhetorical relations, such as Evidence or Result.

I will claim that both approaches to contrast, (A) and (B), capture part of its realization: the underlying semantic relationship required to license the use of contrastive markers often involves focussed referents that can be considered contrastive topics. The greater the degree of abstraction a contrastive topic has, the more context-dependent the information needed to identify it is. Identifying the rhetorical relation that the contrast relation has with the earlier discourse then should aid in identifying these contrastive topics. I additionally hypothesize that in cases where a contrast relation incorporates concrete references to discourse referents that are identifiable as contrastive topics, the contrast relation itself will tend to be related via a Narration relation to the rest of the discourse. For example, in (8) the second sentence contrasting the topics of uneasy sisters vs. a delighted mother relates via Narration to the first:

(8) Jane had not been gone long before it rained hard. Her sisters were on uneasy for her, but her mother was delighted. (Austen, 1813)

Thus heuristics for identifying the more tangible contrastive topics will also aid in developing heuristics for determining how a contrastive relation is rhetorically related to the rest of the discourse. Q2 can best be studied with corpus and experimental research after Q1 has been answered.

Q3 How can the production and interpretation of contrast in dialogue best be modelled formally?

Optimality Theory (OT) has been very successful with phenomena affected by syntactic, semantic and pragmatic information (Hendriks & de Hoop, 2001; de Swart & Zwarts, to appear). Contrast clearly falls into this category and therefore I will use a form of OT as a framework for the analysis. Contrast relations also show speaker-hearer effects. For example:

(9) Lars: I wanted to buy wine this Saturday, but it was after 3PM.

A hearer not familiar with the restrictions on Swedish alcohol sales will be confused by why Lars seems to perceive buying wine on Saturday incongruous with trying to do it after 3PM. If Lars had said and instead of but the hearer would also interpret it as the beginning of a narrative, and that Lars probably did purchase wine. The hearer must reason that the Lars must have had some reason for using but rather than and, so each conjunct must imply some contradiction. The hearer must first, accommodate that buying wine in Sweden after 3PM on Saturday must be impossible, and second, conclude that Lars must have been unsuccessful.

Bidirectional Optimality Theory (Blutner 2000) captures this type of reasoning in a straightforward way. Also, BiOT could be extended to model of speakers in dialogue, a novel use of the framework. I have successfully used BiOT in earlier work (i.e. Spenader & Blutner, to appear; Spenader 2004b; Hendriks & Spenader 2004) and believe it is the framework that best can coherently model the above mentioned characteristics and provide results tha can be easily incorporated into an application.

Approach

Using a data-driven methodology is necessary to answer Q1 and Q2 and this data will also empirically inform any claims made regarding Q3.

Corpora provide the preceding and following contexts necessary to study both the conditions of use, and the effect of a contrast relation. Study of spoken language corpora transcripts, e.g. the spoken part of the British National Corpus and the London-Lund Corpus will be balanced with newspaper texts. Spoken corpora allows access to the reaction of other speakers, extremely helpful in interpreting complex pragmatic phenomena (as evidenced in e.g. Spenader 2002 and Spenader 2003), while newspaper texts are more comparable with other studies of computational discourse. Some collaborative annotation will be done with agreement evaluated with the Kappa statistic (Carletta et al 1998).

However, corpus study seldom provides suitably informative minimal pairs to illustrate specific features. This work will need to be supplemented with constructed examples that can be tested experimentally with judgment and interpretation tasks with online tests as done in Spenader (2004b).

Innovation

First, studying contrast in discourse represents a unique case study of how the two perspectives of rhetorical discourse structure (A) and information structure (B) relate, and will support the efforts of other researchers in determining the correct way to treat coherence-bearing phenomena. Second, this study also represents the first large scale data-driven study of contrast, necessary if contrastive relations are ever to be incorporated in natural language applications. Third, the use of bidirectional optimality theory to model dialogue is a pioneering application of the theory in a new function. Finally, the incorporation of speaker-hearer effects explicitly in the analysis by using BiOT is an original perspective on contrast.

Literature references

Asher,N. (2004). Discourse Topic, Theoretical Linguistics, 30(2-3) 163-201

Asher, N. and A. Lascarides (2003). Logics of Conversation. Cambridge University Press.

Asher, N. and A. Lascarides (1993). Temporal interpretation, discourse relations and commonsense

entailment. Linguistics and Philosophy, 16, 437-493.

Austin, J. (1813). Pride and Prejudice.

Blutner, R. (2000). Some Aspects of Optimality in Natural Language Interpretation, Journal of

Semantics 17, 189-216.

Carletta J., A. Isard, S. Isard, J. Kowtko, G. Doherty-Sneddon & A. Anderson (1997). The

Reliability of a Dialogue Structure Coding Scheme. Computational Linguistics, 23(1),

13- 32.

Carlson, L., D. Marcu, & M. Okurowsky (2001). Building a discourse-tagged corpus in

the framework of rhetorical structure theory. In Proceedings of the 2nd SigDial Workshop on Discourse and Dialogue.

Hendriks, P. & H. de Hoop (2001), Optimality Theoretic Semantics. Linguistics and Philosophy

24(1), 1-32.

Hendriks, P. & J. Spenader (2004). Pronoun Interpretation Problem, ESSLLI 2004

Workshop on Semantic Approaches to Binding Theory (2004). A bidirectional explanation of the pronoun interpretation problem.

Hobbs, J. (1990) Literature and Cognition. CSLI Lecture Notes Number 21.

Kehler, A. (2000). Coherence and the resolution of ellipsis. Linguistics and Philosophy 23,533-575.

Krifka, M. (1999). Additive particles under stress. In SALT 8, Cornell, pp. 111{128. CLC

Publications.

Ogden, C. K. (1932). Opposition: a linguistic and psychological analysis. London.

Maier, E. & J. Spenader (2004). Contrast as denial. In Proceedings of Catalog ‘04, Formal

Semantics and Pragmatics of Discourse, 156-157. Barcelona, Spain,

Mann, W. and S. Thompson (1988). Rhetorical structure theory: Towards a functional theory of text

organization. TEXT, 8(2), 243-281.

Martin, J. (1992). English Text. System and Structure.John Benjamin Publishing Company.

Spenader, J. (2002). Presuppositions in Spoken Discourse. Ph. D. thesis, Stockholm

University.

Spenader, J. (2003). Factive presuppositions, accommodation and information structure.

Journal of Logic, Language and Information 12(3), 351-368.

Spenader, J. (2004a). Modality realization as contrast in discourse. Journal of Semantics

21(3), 113-131.

Spenader, J. (2004b). Optimality Theory and Pragmatics, Chapter Variation in choice of

Swedish Demonstratives, 228-250. Palgrave/Macmillian.

Spenader, J. (2005). Contrast in Context: How do Contrastive Relations Fit with the Rest

of a Discourse? Invited talk at Nijmegen Semantics Colloquium, April 12, 2005, Radboud University Nijmegen.

Spenader, J. & R. Blutner (to appear). The Nature of Systematicity in Natural Language.

Proceedings of the KNAW Colloquium “Cognitive Foundations in Interpretation”, October 27, Amsterdam

Spooren, W. (1989). Some Aspects of the Form and Interpretation of Global Contrastive

Coherence Relations. University of Nijmegen: Sneldruk Enschede.

Stede, M. (2003). Does discourse processing need discourse topics? Theoretical

Linguistics 30, 241- 253.

de Swart, H. & J, Zwarts (to appear). Interpretation as conflict resolution. To appear in a special

issue on semantics of the journal Ilha do desterro. Appeared in the CKI pre-print series as

CKI pre-print no. 045. (May 2004).

Umbach, C. (2001). Contrast and contrastive topic. In I. Kruiff -Kobayana and M.

Steedman (Eds.), Proc. of the ESSLLI Workshop on Information Structure, Discourse Structure and Discourse Semantics, Helsinki, Finland. ESSLLI.

Webber, B. L., M. Stone, A. K. Joshi, & A. Knott (2003). Anaphora and discourse

structure. Computational Linguistics, 29(4), 545-587.

Winter, Y. & M. Rimon (1994). Contrast and implication in natural language. Journal of

Semantics, 11(4), 365-406

Recent Results

Spenader, J. & G. Stulp (2007). Antonymy in Contrast Relations. Seventh International Workshop on Computational Semantics, Tilburg, 10-12 January, 2007.

Lobanova, A. & Spenader, J. (2007). Incorporating Polarity in Lexical Resources. In Proceedings of GL 2007, Approaches to the Generative Lexicon. Paris, May 2007.