Comparing System-response Retrieval Models for Open-domain and Casual Conversational Agent

Abstract : This paper studies corpus-based process to select a system-response usable both in chatterbot or as a fallback strategy. It presents, evaluates and compares two selection methods that retrieve and adapt a system-response from the OpenSubtitles2016 corpus given a human-utterance. A corpus of 800 annotated pairs is constituted. Evaluation consists in objective metrics and subjective annotation based on the validity schema proposed in the RE-WOCHAT shared task. Our study indicates that the task of assessing the validity of a system-response given a human-utterance is subjective to an important extent, and is thus a difficult task. Comparisons show that the selection method based on word embedding performs objectively better than the one based on TF-IDF in terms of response variety and response length.
Liste complète des métadonnées

Cited literature [19 references]  Display  Hide  Download

https://hal.archives-ouvertes.fr/hal-01782262
Contributor : Guillaume Dubuisson Duplessis <>
Submitted on : Tuesday, May 1, 2018 - 2:57:03 PM
Last modification on : Monday, March 18, 2019 - 4:24:29 PM
Document(s) archivé(s) le : Monday, September 24, 2018 - 6:31:51 PM

File

RP-269.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01782262, version 1

Citation

Franck Charras, Guillaume Dubuisson Duplessis, Vincent Letard, Anne-Laure Ligozat, Sophie Rosset. Comparing System-response Retrieval Models for Open-domain and Casual Conversational Agent. Second Workshop on Chatbots and Conversational Agent Technologies (WOCHAT@IVA2016), 2016, Los Angeles, United States. 〈hal-01782262〉

Share

Metrics

Record views

77

Files downloads

67