Improving Context-Awareness on Multi-Turn Dialogue Modeling with Extractive Summarization Techniques

Abstract

The study of context-awareness in multi-turn generation-based dialogue modeling is an important but relatively underexplored topic. Prior research has employed hierarchical structures to enhance the context-awareness of dialogue models. This paper aims to address this issue by utilizing two extractive summarization techniques, namely the PMI topic model and the ORACLE algorithm, to filter out unimportant utterances within a given context. Our proposed approach is assessed on both non-hierarchical and hierarchical models using the distracting test, which evaluates the level of attention given to each utterance. Our proposed methods gain significant improvement over the baselines in the distracting test.