ConvNTM: Conversational Neural Topic Model

Topic models have been thoroughly investigated for multiple years due to their great potential in analyzing and understanding texts. Recently, researchers combine the study of topic with deep learning techniques, known as Neural Models (NTMs). However, existing NTMs are mainly tested based on general document modeling without considering different textual analysis scenarios. We assume that there characteristics model topics tasks. In this paper, we propose a Conversational Model (ConvNTM) designed particular conversational scenario. Unlike modeling, conversation session lasts turns: each short-text utterance complies single distribution these distributions dependent across turns. Moreover, roles conversations, a.k.a., speakers addressees. partially determined by such conversations. take factors into account conversations via multi-turn multi-role formulation. also leverage word co-occurrence relationship new training objective further improve quality. Comprehensive experimental results benchmark datasets demonstrate our proposed ConvNTM achieves best performance both typical downstream tasks within research (i.e., dialogue act classification response generation).