Counterfactual Data Augmentation via Perspective Transition for Open-Domain Dialogues

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 1635 - 1648 December 7-11, 2022 ©2022 Association for Computational Linguistics Counterfactual Data Augmentation via Perspective Transition for Open-Domain Dialogues Jiao Ou1,2, Jinchao Zhang3, Yang Feng1,2∗, Jie Zhou3 1 Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS) 2 University of Chinese Academy of Sciences 3 Pattern Recognition Center, WeChat AI, Tencent Inc, China {oujiao17b,fengyang}@ict.ac.cn, {dayerzhang,withtomzhou}@tencent.com Abstract The construction of open-domain dialogue sys- tems requires high-quality dialogue datasets. The dialogue data admits a wide variety of responses for a given dialogue history, espe- cially responses with different semantics. How- ever, collecting high-quality such a dataset in most scenarios is labor-intensive and time- consuming. In this paper, we propose a data augmentation method to automatically aug- ment high-quality responses with different se- mantics by counterfactual inference. Specifi- cally, given an observed dialogue, our counter- factual generation model first infers semanti- cally different responses by replacing the ob- served reply perspective with substituted ones. Furthermore, our data selection method filters out detrimental augmented responses. Experi- mental results show that our data augmentation method can augment high-quality responses with different semantics for a given dialogue history, and can outperform competitive base- lines on multiple downstream tasks. 1 Introduction Open-domain dialogue systems have attracted much attention (Chen et al., 2017; Huang et al., 2020; Ni et al., 2021; Fu et al., 2022) due to their potential applications. Generally, training open- domain dialogue systems requires high-quality dia- logue datasets. The dialogue data admits a wide va- riety of responses for a given dialogue history (Hou et al., 2018). Specifically, a given dialogue his- tory can exist many valid responses with different semantics, and the response of each semantic in- formation can also have abundant alternative ex- pressions (Li et al., 2019). However, manually col- lecting high-quality such datasets is usually labor- intensive and time-consuming in practice. A feasible solution to address this problem is to use data augmentation techniques. Currently, ∗Work done while Jiao Ou was interning at WeChat AI. Yang Feng is the corresponding author. Our code is public at https://github.com/ictnlp/CAPT. Smoking is harmful for your health. Dialogue history� What in your opinion can be done to stop smoking? Observed response� I have regular check-ups, and do not have any disease. Counterfactual response� stop smoking Observed reply perspective� check-ups What if stop smoking check-ups Substituted reply perspective� Figure 1: An example of a counterfactual response, which is a semantically different response re-inferred by changing the observed reply perspective. some data augmentation methods have been used in open-domain dialogues (Sennrich et al., 2016; Niu and Bansal, 2019; Li et al., 2019; Cai et al., 2020; Zhang et al., 2020a; Xie et al., 2022) to augment data. However, the augmented data have limited semantic differences from the observed data based on the restrained changes. These existing methods only consider word- or sentence-level alternative expressions of the observed data without augment- ing more valid responses with different semantics. …

Content truncated to 500 words