Bayesian Approaches to Preference-based Answer Generation in Dialogue

点击翻译按钮获取中文摘要

This project builds on the results of the DFG-funded project 'Pragmatic Requirements for Answer Generation in a Sales Dialogue'. We intend to investigate the use of Bayesian probabilistic approaches to the generation of pragmatically appropriate answers in a dialogue system. We want to study three different linguistic means that show a sensitivity to interlocutors' (possibly opposing) preferences: numerical and gradable expressions, biscuit conditionals (relevance conditionals), and evaluative adverbials. Each of these means are able to trigger implicatures based on supposed preferences. These phenomena require the dialogue system to manage preferences and pragmatically inferred meaning in its information states. This in turn requires probabilistic reasoning, given that preferences are not always known with high certainty. Data acquisition and formal modeling will be methodologically intertwined. The evaluation of the respective components of the dialogue system will be performed by comparing them with a semantically oriented base system exhibiting no pragmatic competence.