SimOAP: Improve Coherence and Consistency in Persona-based Dialogue Generation via Over-sampling and Post-evaluation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 9945–9959 July 9-14, 2023 ©2023 Association for Computational Linguistics SimOAP: Improve Coherence and Consistency in Persona-based Dialogue Generation via Over-sampling and Post-evaluation Junkai Zhou1,2, Liang Pang1∗, Huawei Shen1,2, Xueqi Cheng1,2∗ 1Data Intelligence System Research Center, Institute of Computing Technology, CAS 2University of Chinese Academy of Sciences {zhoujunkai20z, pangliang,shenhuawei,cxq}@ict.ac.cn Abstract Language models trained on large-scale cor- pora can generate remarkably fluent results in open-domain dialogue. However, for the persona-based dialogue generation task, consis- tency and coherence are also key factors, which are great challenges for language models. Ex- isting works mainly focus on valuable data fil- tering, model structure modifying, or objective function designing, while their improvements are limited and hard to generalize to all types of pre-trained language models. However, we find that language models can produce consistent and coherent responses if we consider enough generations. Thus, the problems lay in large- scale response generation and target response selection. In this work, a simple but effec- tive two-stage SimOAP strategy is proposed, i.e., over-sampling and post-evaluation. The over-sampling stage takes large-scale responses from existing trained models efficiently via off- the-shelf distilling and compressing methods, and the post-evaluation stage selects a good re- sponse based on multiple well-designed evalua- tion metrics from large-scale candidates. Exper- imental results show that the proposed plug-in SimOAP strategy improves the backbone mod- els and outperforms the baseline strategies in both automatic and human evaluations. 1 Introduction Open-domain dialogue systems need to give ap- propriate responses based on history utterances. An ideal open-domain dialogue system should gen- erate consistent, coherent, and diverse responses. Part of the existing open-domain dialogue gener- ation work focuses on improving the diversity of responses (Wang et al., 2021), while avoiding gen- erating generic responses and achieving good re- sults. How to improve the consistency of dialogue generation is also an urgent problem to be solved (Kim et al., 2020). In addition, there is still the ∗Corresponding authors 0 (0.00) 2000 (1.48) 4000 (1.74) 6000 (2.07) 8000 (2.67) 10000 (553.35) The rank of the selected response in all responses (PPL) 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 Ratio of good responses 1e−3 Figure 1: We use a dialogue model to generate 10,000 responses each for 100 utterances in PersonaChat and use perplexity (PPL) to rerank them. The response is good when the TF-IDF similarity between the response and ground truth is above 0.25 and the result of the natural language inference model between the response and persona information is entailment. PPL in brackets is the average value of all responses in each rank. problem of poor coherence in dialogue generation (Ye et al., 2021). To improve the consistency and coherence of dialogue generation, the existing works mainly im- prove from three aspects: valuable data construc- tion and filtering (Zhang et al., 2018; Song et al., 2020b), model structure modifying (Song et al., 2021; Zou et al., 2021; Liu et al., 2023) and objec- …

Content truncated to 500 words