Modeling Multi-turn Conversation with Deep Utterance Aggregation

TL;DR

This paper introduces a deep utterance aggregation model for multi-turn conversation response matching, effectively capturing interactions among previous utterances to improve dialogue system performance.

Contribution

It proposes a novel deep utterance aggregation approach with self-matching attention for better context modeling in multi-turn conversations.

Findings

Outperforms state-of-the-art methods on three benchmarks

Achieves superior response matching accuracy

Effective on a new e-commerce dialogue corpus

Abstract

Multi-turn conversation understanding is a major challenge for building intelligent dialogue systems. This work focuses on retrieval-based response matching for multi-turn conversation whose related work simply concatenates the conversation utterances, ignoring the interactions among previous utterances for context modeling. In this paper, we formulate previous utterances into context using a proposed deep utterance aggregation model to form a fine-grained context representation. In detail, a self-matching attention is first introduced to route the vital information in each utterance. Then the model matches a response with each refined utterance and the final matching score is obtained after attentive turns aggregation. Experimental results show our model outperforms the state-of-the-art methods on three multi-turn conversation benchmarks, including a newly introduced e-commerce dialogue corpus.