Memory-based model with multiple attentions for multi-turn response selection

  • Xingwu Lu
  • , Man Lan*
  • , Yuanbin Wu
  • *Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

Abstract

In this paper, we study the task of multi-turn response selection in retrieval-based dialogue systems. Previous approaches focus on matching response with utterances in the context to distill important matching information, and modeling sequential relationship among utterances. This kind of approaches do not take into account the position relationship and inner semantic relevance between utterances and query (i.e., the last utterance). We propose a memory-based network (MBN) to build the effective memory integrating position relationship and inner semantic relevance between utterances and query. Then we adopt multiple attentions on the memory to learn representations of context with multiple levels, which is similar to the behavior of human that repetitively think before response. Experimental results on a public data set for multi-turn response selection show the effectiveness of our MBN model.

Original languageEnglish
Title of host publicationNeural Information Processing - 25th International Conference, ICONIP 2018, Proceedings
EditorsAndrew Chi Sing Leung, Seiichi Ozawa, Long Cheng
PublisherSpringer Verlag
Pages296-307
Number of pages12
ISBN (Print)9783030041786
DOIs
StatePublished - 2018
Event25th International Conference on Neural Information Processing, ICONIP 2018 - Siem Reap, Cambodia
Duration: 13 Dec 201816 Dec 2018

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11302 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference25th International Conference on Neural Information Processing, ICONIP 2018
Country/TerritoryCambodia
CitySiem Reap
Period13/12/1816/12/18

Keywords

  • Memory network
  • Multi-turn conversation
  • Neural networks
  • Response selection

Fingerprint

Dive into the research topics of 'Memory-based model with multiple attentions for multi-turn response selection'. Together they form a unique fingerprint.

Cite this