|
|
"RESOLVING FRAGMENTS: A THEORY
OF CONTEXT FOR CONVERSATION"
Jonathan Ginzburg King's College London Wednesday, May 16, 12:00noon, MJH Rm 126 Non sentential utterances (NSUs), examples of which are in (1)-(4), are pervasive in conversation: (1) A: Did Bo leave? B: Yes / No, Jo/ Bo? (2) A: Did Bo ... B: leave? (3) A: Did Bo ... No, Jo leave? (4) A: [in a ticket office] A return to Maidenhead please. NSUs have had a rather mixed reputation in theoretical and computational linguistics. Beliefs about the intrinsic messiness of conversation, in particular its being littered with fragments, have been used as important motivation for a strong modularity assumption in language acquisition. Conversely, computational linguists have argued that fragment resolution generally requires plan recognition techniques (e.g. Allen and Perrault, 1980, Carberry, 1991). In the first part of this talk I will briefly survey some recent work, jointly with Raquel Fernandez and Shalom Lappin, showing that conversational NSUs are actually amenable to domain independent classification: a taxonomy of fewer than 20 classes reliably classifies the NSUs in the British National Corpus and the taxonomy can be learnt by a variety of machine learning algorithms. The main part of the talk will address the issue of how to develop a theory of context that is capable of underpinning a grammatical analysis of the wide range of observed fragments, as exemplified above. My main claim will be that given a sufficiently detailed theory of conversational interaction, fragment resolution is relatively straightforward, akin in a sense to the resolution of traditional indexical elements. The resulting theory of context allows us to associate a hierarchy of complexity among NSUs. Some predictions this makes about the order of acquisition of NSUs, in contrast with traditional, syntactically oriented views of resolution, will be discussed. 
|