Response Selection for Multi-Party Conversations with Dynamic Topic Tracking

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 6581–6591, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 6581 Response Selection for Multi-Party Conversations with Dynamic Topic Tracking Weishi Wang§‡, Steven C.H. Hoi§, Shafiq Joty§‡ §Salesforce Research Asia ‡Nanyang Technological University, Singapore §{weishi.wang,shoi,sjoty}@salesforce.com Abstract While participants in a multi-party multi-turn conversation simultaneously engage in multi- ple conversation topics, existing response se- lection methods are developed mainly focus- ing on a two-party single-conversation sce- nario. Hence, the prolongation and transition of conversation topics are ignored by current methods. In this work, we frame response selection as a dynamic topic tracking task to match the topic between the response and rel- evant conversation context. With this new for- mulation, we propose a novel multi-task learn- ing framework that supports efficient encoding through large pretrained models with only two utterances at once to perform dynamic topic disentanglement and response selection. We also propose Topic-BERT an essential pretrain- ing step to embed topic information into BERT with self-supervised learning. Experimental results on the DSTC-8 Ubuntu IRC dataset show state-of-the-art results in response selec- tion and topic disentanglement tasks outper- forming existing methods by a good margin.1 1 Introduction In recent years, with the influx of deep learning methods in natural language processing (NLP), there has been a lot of interests in building effective task-oriented dialogue systems that can assist peo- ple in real-world business such as booking tickets, ordering food and solving technical issues (Bui, 2006). Retrieval-based response generation that selects a suitable response from a pool of candi- dates (pre-existing human responses) has become a popular approach to framing dialog. Compared to the generation-based systems that generate novel utterances (Serban et al., 2016), retrieval-based sys- tems produce fluent, grammatical and informative responses (Weston et al., 2018; Henderson et al., 1Code will be available at https://github.com/ salesforce/TopicBERT. <_timello> sorry, but I lost my link, repeating the question: anybody knows why I can't play any .mpg, etc? it shows me the sound, but not shows me the screen <Nafallo> _timello: probably missing codecs <Nafallo> hmm, anyoneelse got troubles with docbook-dssslversion 1.78-4? <danhunt> Check http://www.desktopos.com/reviews.php?op=PrintReview&id=21 for .mpg tips. <Nafallo> it works now :-P, takes a bit more to downgrade through aptitude than upgrade ;-) <_timello> danhunt, I installed mplayerand the essential codecs package, but it still isn't working. I didn't find why <Nafallo> Well can I move the drives <danhunt> Nafallo: you can?t move the drives, definitely not. This is the problem with RAID : ) <Nafallo> danhunt: haha, yeah <Nafallo> _timello: run mplayerfrom a terminal and check the output? Figure 1: A (truncated) multi-party conversation from Ubuntu IRC log. Curved arrows show the ‘reply-to’ links between utterances. We use different colors to represent different conversation topic clusters. 2019). Also compared to the traditional modular approach, it does not rely on dedicated modules for language understanding, dialog management, and generation, thus simplifying the system design. Due to these reasons, retrieval-based systems have been widely adopted in commercial dialogue sys- tems (Gao et …

Content truncated to 500 words