Multi-turn Response Selection using Dialogue Dependency Relations

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 1911–1920, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 1911 Multi-turn Response Selection using Dialogue Dependency Relations Qi Jia1 Yizhu Liu2 Siyu Ren3 Kenny Q. Zhu4∗ Shanghai Jiao Tong University Shanghai, China {1Jia qi,2liuyizhu,3roy0702}@sjtu.edu.cn 4kzhu@cs.sjtu.edu.cn Haifeng Tang China Merchants Bank Credit Card Center Shanghai, China thfeng@cmbchina.com Abstract Multi-turn response selection is a task de- signed for developing dialogue agents. The performance on this task has a remarkable im- provement with pre-trained language model- s. However, these models simply concatenate the turns in dialogue history as the input and largely ignore the dependencies between the turns. In this paper, we propose a dialogue extraction algorithm to transform a dialogue history into threads based on their dependen- cy relations. Each thread can be regarded as a self-contained sub-dialogue. We also pro- pose Thread-Encoder model to encode thread- s and candidates into compact representations by pre-trained Transformers and finally get the matching score through an attention layer. The experiments show that dependency relations are helpful for dialogue context understanding, and our model outperforms the state-of-the-art baselines on both DSTC7 and DSTC8*, with competitive results on UbuntuV2. 1 Introduction Dialogue system is an important interface between machine and human. An intelligent dialogue agent is not only required to give the appropriate response based on the current utterance from the user, but also consider the dialogue history. Dialogue con- text modeling has been a key point for developing such dialogue systems, including researches on s- tate tracking (Eric et al., 2019; Ren et al., 2019), topic segmentation (Nan et al., 2019; Kim, 2019), multi-turn response selection (Tao et al., 2019; Gu et al., 2019), next utterance generation (Zhang et al., 2019; Chen et al., 2019), etc. In this paper, we tar- get on the multi-turn response selection task, which is first proposed by Lowe et al. (2015) and is also a track in both DSTC7 (Gunasekara et al., 2019) and DSTC8 (Kim et al., 2019). ∗ The corresponding author. Dialogue History Response: B: Those are all the services that load on startup. A: How can i speed up ubuntu's initialization process ? C: My grub menu not displayed while starting ubuntu, help! B: Stop all the services you don’t need from loading on startup. C: The services are listed in /etc/rc2. B: Comment out the vars in /etc/default/grub … A: I have no idea what those files are for. C: Ok, that worked, thanks a lot! Figure 1: An example of the tangled dialogue history. A, B and C are three participants. Texts in different colors represent different dialogue threads. Given a dialogue history made up of more than one utterance, the selection task is to choose the most possible next utterance from a set of candidate responses. Previous work on this task can be rough- ly divided into two categories: sequential models and hierarchical models. The former ones, includ- ing (Lowe et al., 2015; Yan et al., 2016; Chen and Wang, …

Content truncated to 500 words