Fine-grained Post-training for Improving Retrieval-based Dialogue Systems

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1549–1558 June 6–11, 2021. ©2021 Association for Computational Linguistics 1549 Fine-grained Post-training for Improving Retrieval-based Dialogue Systems Janghoon Han1,3, Taesuk Hong1, Byoungjae Kim1, Youngjoong Ko2, Jungyun Seo1 1Department of Computer Science and Engineering, Sogang University 2Department of Computer Science and Engineering, Sungkyunkwan University 3LG AI Research janghoon.han@lgresearch.ai {hongtaesuk,wiz3021,seojy}@sogang.ac.kr yjko@skku.edu Abstract Retrieval-based dialogue systems display an outstanding performance when pre-trained lan- guage models are used, which includes bidirec- tional encoder representations from transform- ers (BERT). During the multi-turn response se- lection, BERT focuses on training the relation- ship between the context with multiple utter- ances and the response. However, this method of training is insufficient when considering the relations between each utterance in the context. This leads to a problem of not completely un- derstanding the context flow that is required to select a response. To address this issue, we propose a new fine-grained post-training method that reflects the characteristics of the multi-turn dialogue. Specifically, the model learns the utterance level interactions by train- ing every short context-response pair in a di- alogue session. Furthermore, by using a new training objective, the utterance relevance clas- sification, the model understands the seman- tic relevance and coherence between the di- alogue utterances. Experimental results show that our model achieves new state-of-the-art with significant margins on three benchmark datasets. This suggests that the fine-grained post-training method is highly effective for the response selection task.1 1 Introduction Constructing a dialogue system that can natu- rally and consistently interact with humans is cur- rently a popular research topic. There are two ap- proaches for the implementation of a dialogue sys- tem: generation-based and retrieval-based methods. The latter approach aims to select the correct re- sponse among the response candidates. In the initial multi-turn response selection, Lowe et al. (2015) proposed leveraging RNN to match the dialogue context with a response. Later, with the advent of 1https://github.com/hanjanghoon/BERT_ FP the attention mechanism (Bahdanau et al., 2015; Luong et al., 2015; Vaswani et al., 2017), multi- turn response selection models that use the atten- tion mechanism have been proposed (Zhou et al., 2018). Recently, pre-trained language models, such as bidirectional encoder representations from trans- formers, BERT, have been applied to a variety of response selection models (Vig and Ramea, 2019; Lu et al., 2020; Gu et al., 2020), and they have shown excellent performance. Recently, pre-trained language models have been widely used in several natural language processing areas, such as question answering and dialogue sys- tems. One of the best pre-trained language models, BERT (Devlin et al., 2019), is initially pre-trained on a large and general domain corpus, and then it is fine-tuned to adapt to specific tasks. Since BERT is pre-trained with general data, its perfor- mance can be improved by post-training to adapt to domain-specific data. Some previous studies (Xu et al., 2019; Whang et al., 2020) proposed a post- training method that can learn domain data …

Content truncated to 500 words