Chapter 6: Discourse and Dialogue
Barbara Grosz
Harvard University, Cambridge, Massachusetts, USA
The problems addressed in discourse research aim to answer two general kinds of questions: (1) what information is contained in extended sequences of utterances that goes beyond the meaning of the individual utterances themselves? (2) how does the context in which an utterance is used affect the meaning of the individual utterances, or parts of them?
Computational work in discourse has focused on two different types of discourse: extended texts and dialogues, both spoken and written. Although there are clear overlaps between these---dialogues contain text-like sequences spoken by a single individual and texts may contain dialogues---the current state of the art leads research to focus on different questions for each. In addition, application opportunities and needs are different. Work on text is of direct relevance to document analysis and retrieval applications whereas work on dialogue is of import for human-computer interfaces regardless of the modality of interaction. A good sense of the current state of research in text interpretation can be gained from reading the papers on text interpretation in a recent special issue of Artificial Intelligence (hereafter, AIJ-SI), [HSAM94,JR94,PPWF94].
Text and dialogue have, however, two significant commonalities.
First, a major result of early work in discourse was the determination
that discourses divide into discourse segments much
like sentences divide into phrases. Utterances group into segments,
with the meaning of a segment encompassing more than the meaning of
the individual parts. Different theories vary on the factors they
consider central to explaining this segmentation; a review of the
alternatives can be found in a previous survey [GPS89]
(hereafter, Discourse Survey).
However, many of the
implications for language processing are shared. For example,
segment boundaries need to be detected; recent work suggests
there are intonational indicators of these boundaries in
spoken language (e.g., [GH92] and the references cited
in this paper) and can be used to improve speech synthesis
(e.g., [DH88]).
Second, discourse research on the interpretation of referring expressions, including pronouns and definite descriptions (e.g., le petit chat, das grüne Buch), and the event reference aspect of verb phrase interpretation (e.g., the relationship between the buying and arriving events in the sequence John went to Mary's house; he had bought flowers at her favorite florist's also is relevant to both text and dialogue. Work on these problems before 1990 is described in Discourse Survey.
The major lines of research on determining what information a discourse carries beyond what is literally expressed in the individual sentences the discourse comprises fall into two categories which, following Hobbs, we will refer to as informational and intentional. There are currently efforts to combine these two approaches (e.g., [Keh94,Keh95,MP92]); this is an important area of research.
According to the informational approaches, the coherence of discourse follows from semantic relationships between the information conveyed by successive utterances. As a result, the major computational tools used here are inference and abduction on representations of the propositional content of utterances. Discourse Survey describes work in this area under inference-based approaches; more recent work in this area is presented in AIJ-SI.
According to the intentional approaches the coherence of
discourse derives from the intentions of speakers and writers, and
understanding depends on recognition of those intentions. Thus,
these approaches follow [Gri69]; early work in this area
drew on speech act theory [Sea69]. A major
insight of work in this area was to recognize the usefulness of
applying AI planning techniques; this work is described in
Discourse Survey. Recently various limitations of this approach
have been recognized. In particular, as originally argued by
[Sea90] and [GS90], models of
individual plans are not adequate for understanding
discourse; models of collaborative plans or joint intentions are required. A variety of approaches to developing such
models are currently underway
[GK93,STW
94,CL90] and used
for dialogue [Loc93,Loc94,Loc95].
Research in this area also splits into two approaches, those that examine the interaction of choice or interpretation of expression with focus of attention, and those that are coherence-based.
Focus of attention interacts with the interpretation and generation of pronouns and definite descriptions [GS86]. The coherence-based approaches have been taken with the informational approaches described above. The main new issues in this area concern how to combine these approaches as it is clear that both kinds of consideration play roles both in determining which expressions to use and how to interpret expressions in context. The focus-based approaches have been applied cross-linguistically;
because this is a cognitively-oriented approach it should have application to multi-media interfaces even when natural language is not being used, or when only a restricted subset can be handled.