Abstract

This dissertation provides a computational model for generating and interpreting indirect answers. An indirect answer conversationally implicates the requested information, i.e., it licenses a pragmatic inference about the intended meaning of the response. In my model, pragmatic knowledge about full answers, consisting of direct answers and extra appropriate information, is encoded in discourse plan operators. In appropriate contexts, the extra information in a full answer may be used as an indirect answer. ;Interpretation is modeled as discourse plan recognition. The inputs to interpretation include the discourse expectation that the speaker S will inform the hearer H if the questioned proposition is true, H's beliefs, the semantic interpretation of the utterances given by S during S's turn, and pragmatic knowledge. The output is a ranked set of discourse plans which it is plausible to ascribe to S. During discourse plan recognition, logical inference plays a role: in verifying that the applicability conditions and goals of a candidate plan are consistent with H's beliefs about S's beliefs and goals, in proving that the coherence, relations specified by the plan are plausible, and in hypothesizing missing parts of the plan based upon assumptions about discourse coherence. ;The inputs to generation include the same discourse expectation as above, S's beliefs, and pragmatic knowledge. The output is a discourse plan. The plan is constructed in two steps: content planning and plan pruning. Content planning is performed by top-down construction of a hierarchical discourse plan for a full answer. S's motivation for including subactions providing extra information is modeled by means of stimulus conditions which reflect general concerns of politeness and cooperativity. During plan pruning, the generator determines what parts of the full answer are inferable and thus do not need to be given explicitly by S. ;This model handles a wider variety of types of indirect answers than previous models, provides new techniques for discourse plan inference and content planning, and provides aIl approach for other types of conversational implicature.