Skip to main navigation Skip to search Skip to main content

Learning to start for sequence to sequence based response generation

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

Abstract

Response Generation which is a crucial component of a dialogue system can be modeled using the Sequence to Sequence (Seq2Seq) architecture. However, this kind of method suffers from vague responses of little meaningful content. One possible reason for generating vague responses is the different distribution of the first word between the generated responses and human responses. In fact, the Seq2Seq based method tends to generate high-frequency words in the beginning, which influences the following prediction resulting in vague responses. In this paper, we proposed a novel approach, namely learning to start (LTS), to learn how to generate the first word in the sequence to sequence architecture for response generation. Experimental results show that the proposed LTS model can enhance the performance of the start-of-the-art Seq2Seq model as well as other Seq2Seq models for response generation of short text conversation.

Original languageEnglish
Title of host publicationInformation Retrieval - 24th China Conference, CCIR 2018, Proceedings
EditorsXianxian Li, Chenliang Li, Tie-Yan Liu, Jiafeng Guo, Shichao Zhang
PublisherSpringer Verlag
Pages274-285
Number of pages12
ISBN (Print)9783030010119
DOIs
StatePublished - 2018
Event24th China Conference on Information Retrieval, CCIR 2018 - Guilin, China
Duration: 27 Sep 201829 Sep 2018

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume11168 LNCS
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Conference

Conference24th China Conference on Information Retrieval, CCIR 2018
Country/TerritoryChina
CityGuilin
Period27/09/1829/09/18

Keywords

  • Learning to start
  • Response generation
  • Sequence to sequence

Fingerprint

Dive into the research topics of 'Learning to start for sequence to sequence based response generation'. Together they form a unique fingerprint.

Cite this