GECOR: An End-to-End Generative Ellipsis and Co-reference Resolution Model for Task-Oriented Dialogue

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 4547–4557, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 4547 GECOR: An End-to-End Generative Ellipsis and Co-reference Resolution Model for Task-Oriented Dialogue Jun Quan‡∗, Deyi Xiong‡†, Bonnie Webber§ and Changjian Hu¶ ‡ School of Computer Science and Technology, Soochow University, Suzhou, China § University of Edinburgh, Edinburgh, UK ¶Lenovo Research AI Lab, Beijing, China terryqj0107@gmail.com, bonnie.webber@ed.ac.uk,dyxiong@suda.edu.cn, hucj1@lenovo.com Abstract Ellipsis and co-reference are common and ubiquitous especially in multi-turn dialogues. In this paper, we treat the resolution of el- lipsis and co-reference in dialogue as a prob- lem of generating omitted or referred expres- sions from the dialogue context. We there- fore propose a unified end-to-end Generative Ellipsis and CO-reference Resolution model (GECOR) in the context of dialogue. The model can generate a new pragmatically com- plete user utterance by alternating the gen- eration and copy mode for each user utter- ance. A multi-task learning framework is fur- ther proposed to integrate the GECOR into an end-to-end task-oriented dialogue. In or- der to train both the GECOR and the multi- task learning framework, we manually con- struct a new dataset on the basis of the pub- lic dataset CamRest676 with both ellipsis and co-reference annotation. On this dataset, in- trinsic evaluations on the resolution of ellip- sis and co-reference show that the GECOR model significantly outperforms the sequence- to-sequence (seq2seq) baseline model in terms of EM, BLEU and F1 while extrinsic evalua- tions on the downstream dialogue task demon- strate that our multi-task learning framework with GECOR achieves a higher success rate of task completion than TSCP, a state-of-the-art end-to-end task-oriented dialogue model (Lei et al., 2018). 1 Introduction Due to the rhetorical principle of saving words and avoiding repetitions, ellipsis and co-reference oc- cur frequently in multi-turn dialogues leaving ut- terances paragmatically incomplete if they are sep- arate from context. Humans can easily understand utterances with anaphorically referenced or absent ∗ Work performed during an internship at Lenovo Re- search AI Lab. †Corresponding author information (e.g., Q2 and Q3 in Table 1) based on the dialogue context while dialogue systems often fail to understand such utterances correctly, which may result in false or incoherent responses. If user utterances can be automatically supple- mented with information that is left out or substi- tuted by anaphora according to the dialogue con- text as humans do (e.g., Q2: I want cheap Italian restaurants. Q3: Yes, I would like the phone num- ber please.), dialogue models may understand user requests correctly and would not generate wrong responses caused by ellipsis and co-reference phe- nomena. Especially in task-oriented dialogue sys- tems, explicitly providing such information to the models can effectively improve the success rate of task completion. In order to achieve this goal, we propose an end- to-end generative ellipsis and co-reference res- olution model (GECOR) for task-oriented dia- logue in this paper. The essential idea behind GECOR is …

Content truncated to 500 words