Mining Clues from Incomplete Utterance: A Query-enhanced Network for Incomplete Utterance Rewriting

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4839 - 4847 July 10-15, 2022 ©2022 Association for Computational Linguistics Mining Clues from Incomplete Utterance: A Query-enhanced Network for Incomplete Utterance Rewriting Shuzheng Si1,2∗, Shuang Zeng1,2∗, Baobao Chang1† 1Key Laboratory of Computational Linguistics, Peking University, MOE, China 2School of Software and Microelectronics, Peking University, China sishuzheng@stu.pku.edu.cn {zengs,chbb}@pku.edu.cn Abstract Incomplete utterance rewriting has recently raised wide attention. However, previous works do not consider the semantic struc- tural information between incomplete utter- ance and rewritten utterance or model the se- mantic structure implicitly and insufficiently. To address this problem, we propose a QUEry- Enhanced Network (QUEEN). Firstly, our proposed query template explicitly brings guided semantic structural knowledge be- tween the incomplete utterance and the rewrit- ten utterance making model perceive where to refer back to or recover omitted tokens. Then, we adopt a fast and effective edit operation scoring network to model the relation between two tokens. Benefiting from extra informa- tion and the well-designed network, QUEEN achieves state-of-the-art performance on sev- eral public datasets. 1 Introduction Multi-turn dialogue modeling, a classic research topic in the field of human-machine interaction, serves as an important application area for prag- matics (Leech, 2003) and Turing Test. The ma- jor challenge in this task is that interlocutors tend to use incomplete utterances for brevity, such as referring back to (i.e., coreference) or omitting (i.e., ellipsis) entities or concepts that appear in di- alogue history. As shown in Table 1, the incom- plete utterance u3 refers to “Smith” (‘‘史密斯”) from u1 and u2 using a pronoun “He” (‘‘他”) and omits “the type of cuisine” (‘‘菜肴的类型”) from u2. This may cause referential ambiguity and se- mantic incompleteness problems if we only read this single utterance u3, which is a common case of downstream applications like retrieval-based di- alogue systems (Boussaha et al., 2019). Moreover, previous studies (Su et al., 2019; Pan et al., 2019) *Equal contribution. †Corresponding author. Turns Utterances with Translation u1 Smith needs to find an expensive restaurant nearby. 史密斯需要在附近找一家昂贵的餐馆。 u2 Does Smith care the type of cuisine? 史密斯关心菜肴的类型吗？ u3 No, he does not care. 不，他不关心。 u 3′ No, Smith does not care about the type of cuisine. 不，史密斯不关心菜肴的类型。 Table 1: An example in multi-turn dialogue including dialogue utterance history u1 and u2, incomplete utter- ance u3 and rewritten utterance u′ 3. also find that coreference and ellipsis exist in more than 70% of the utterances, especially in pro-drop languages like Chinese. These phenomena make it imperative to effectively model dialogue in incom- plete utterance scenarios. To cope with this problem, previous works (Ku- mar and Joshi, 2016a; Elgohary et al., 2019; Su et al., 2019) propose the Incomplete Utterance Rewriting (IUR) task. It aims to rewrite an in- complete utterance into a semantically equivalent but self-contained utterance by mining semantic clues from the dialogue history. Then the gener- ated utterance can be understood without reading dialogue history. For example, in Table 1, after …

Content truncated to 500 words