Attention Based Dialogue Context Selection Model

Weidi Xu, Yong Ren, Ying Tan

Abstract: The particular phenomena of Information Overload and Conversational Dependency in multi-turn dialogues have brought massive noise for feature learning in existing deep learning models. To solve the problem, the Attention Based Dialogue Context Selection Model (ABDCS) is proposed in this paper. This model uses attention mechanism to extract the relationship between current response utterance and previous utterances. Qualitative and quantitative analysis show that ABDCS is able to choose the semantically related utterances in its dialogue history as context and be robust against the noise.