Improving Open-Domain Dialogue Systems via Multi-Turn Incomplete Utterance Restoration

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1824–1833, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1824 Improving Open-Domain Dialogue Systems via Multi-Turn Incomplete Utterance Restoration Zhufeng Pan1,2∗ Kun Bai2∗ Yan Wang2 † Lianqiang Zhou2 Xiaojiang Liu2 1University of California, Los Angeles 2Tencent AI Lab panzhufeng@cs.ucla.edu {bookerbai, brandenwang, tomcatzhou, kieranliu}@tencent.com Abstract In multi-turn dialogue, utterances do not al- ways take the full form of sentences. These incomplete utterances will greatly reduce the performance of open-domain dialogue sys- tems. Restoring more incomplete utterances from context could potentially help the sys- tems generate more relevant responses. To facilitate the study of incomplete utterance restoration for open-domain dialogue systems, a large-scale multi-turn dataset Restoration- 200K1 is collected and manually labeled with the explicit relation between an utterance and its context. We also propose a “pick-and- combine” model to restore the incomplete ut- terance from its context. Experimental results demonstrate that the annotated dataset and the proposed approach significantly boost the re- sponse quality of both single-turn and multi- turn dialogue systems. 1 Introduction Dialogue systems have attracted increasing atten- tion due to the promising potentials on applica- tions like virtual assistants or customer support systems (Hauswald et al., 2015; Poulami Debnath, 2018). However, studies (Carbonell, 1983) show that users of dialogue systems tend to use suc- cinct language which often omits entities or con- cepts made in previous utterances. (also known as non-sentential utterances, (Fern´andez et al., 2005, 2007)). To make appropriate responses, dialogue systems must be equipped with the ability to un- derstand these incomplete utterances. Take Example 1 in Table 1 for instance, con- tents in parentheses are information omitted in the utterance. Humans are capable of comprehending ∗Both authors contributed equally to the work. The work was conducted when Zhufeng Pan was interning at Tencent AI Lab. †Corresponding author 1The dataset is available at: https://ai.tencent. com/ailab/nlp/dialogue/datasets/ those incomplete utterances based on previous ut- terances. For example, A3 means what kind of dessert matches B’s taste, instead of what kind of shop B likes. Failing to understand this utterance would be a disaster for dialogue systems to gen- erate a relevant and coherent response. According to our survey (details in Table 2), in about 60% conversations, fully comprehending current utter- ance depends on previous context. We will refer conversation history (A1 to B2 in above example) as previous utterances, the utterance to be restored (A3) as original utterance, and the complete form of A3 as restored utterance. Studies show that restoring the incomplete questions could help question-answering systems better understand users’ intention (Raghu et al., 2015; Kumar and Joshi, 2017). It inspires us to improve the performance of open-domain di- alogue systems via incomplete utterance restora- tion. However, most existing multi-turn dialogue datasets only provide sets of utterances, with- out any information about relations between utter- ances. In other words, they lack the necessary su- pervisions to restore the incomplete utterance. To …

Content truncated to 500 words