Computational Linguistics Colloquium

Friday May 11, 16:15, Seminar Room, Building 17
NOTE UNUSUAL DAY

Imperfect Communication: Processing and Grammar

One of the important shifts in the study of formal semantics of dialogue from earlier work in formal semantics of natural language concerns abandoning of the assumption of perfect communication. Formulated positively, this involves recognizing that a given conversational participant's understanding another interlocuter's utterance cannot be taken for granted. The existence of such a process of negotiation of whether shared understanding exists (known as grounding, following Clark) has been the subject of much psychological experimentation. There has been some work concerning computational modelling of this process (see e.g. Traum 1994, and subsequent work in the TRINDI project). However, to date there has been relatively little work on providing a formal characterization of utterance understanding and the various ways in which it might fail to arise. One of the central questions one is forced to consider is: what is the nature of the entity on which this reasoning process takes place. Although meanings (in the sense of Montague 1968 et seq -- a function/abstract from contexts to contents) -- might seem a plausible candidate, I will show that such entities are inadequate. One argument for this derives from a class of elliptical utterances used to signal lack of understanding:

A: Did Bo finagle a raise?
B: (i) Bo? / (ii) finagle?

Such utterances are pervasive in conversation and have a variety of possible understandings, that include the following:

1. Check reading: Are you asking if BO (of all people) finagled a raise/Bo FINAGLED a raise (of all actions)?
2. Content identification reading: Who is Bo?/What does it mean to finagle?
3. Lexical identification reading: What word did you mean to pronounce by uttering the sound ``Bo''/``finagle''?

The analysis of such utterances turns out to be beyond standard techniques used in anaphora and ellipsis resolution. I will offer a computational analysis of the resolution of ellipsis in cases such as the above, showing that it requires operations on highly structured, linguistically heterogeneous representations. I characterize these operations and the representations on which they operate and offer an analysis couched in a version of HPSG combined with a theory of information states in dialogue. I sketch an algorithm for the process of utterance integration by conversationalists which leads to grounding or clarification. If I have time, I will also briefly discuss the connection this issue has with the long standing problem of psychological attitude reports.

(Joint work with Robin Cooper, Gothenburg)

If you would like to meet with the speaker, please contact CJ Rupp.