A Model-agnostic Data Manipulation Method for Persona-based Dialogue Generation

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 7984 - 8002 May 22-27, 2022 c⃝2022 Association for Computational Linguistics A Model-Agnostic Data Manipulation Method for Persona-based Dialogue Generation Yu Cao1∗, Wei Bi2†, Meng Fang3, Shuming Shi2, Dacheng Tao1,4 1School of Computer Science, The University of Sydney, Australia 2Tencent AI Lab, Shenzhen, China 3Eindhoven University of Technology (TU/e), Eindhoven, The Netherlands 4JD Explore Academy, Beijing, China ycao8647@sydney.edu.au, victoriabi@tencent.com, m.fang@tue.nl, shumingshi@tencent.com, dacheng.tao@gmail.com Abstract Towards building intelligent dialogue agents, there has been a growing interest in intro- ducing explicit personas in generation models. However, with limited persona-based dialogue data at hand, it may be difficult to train a dia- logue generation model well. We point out that the data challenges of this generation task lie in two aspects: first, it is expensive to scale up current persona-based dialogue datasets; sec- ond, each data sample in this task is more com- plex to learn with than conventional dialogue data. To alleviate the above data issues, we propose a data manipulation method, which is model-agnostic to be packed with any persona- based dialogue generation model to improve its performance. The original training sam- ples will first be distilled and thus expected to be fitted more easily. Next, we show various effective ways that can diversify such easier distilled data. A given base model will then be trained via the constructed data curricula, i.e. first on augmented distilled samples and then on original ones. Experiments illustrate the superiority of our method with two strong base dialogue models (Transformer encoder- decoder and GPT2). 1 Introduction The ability to generate responses with consistent personas is important towards building intelligent dialogue agents. In past years, there has been a growing interest in introducing explicit personas in dialogue generation models (Song et al., 2019; Wolf et al., 2019). A piece of persona text gener- ally consists of profiles and background personal facts. A clipped persona-based dialogue from the PersonaChat (Zhang et al., 2018a) dataset is shown in Figure 1, which covers rich persona features. For ∗Work was done when Yu Cao was an intern at Tencent AI Lab. † Corresponding author Personas of A i work long hours as a veteri so i am Personas of great . i was just reading a book and Very busy , having a large family of five and a husband . I bet you keep busy … … … … want to take a break. taking a break too. you keep busy. Very busy 1.i work as a veterinarian . 2.i am married and have work long hours as of five and a husband . five children . 3.i am a vegetarian . 4.my favorite music is hip hop. work long hours as a veterinarian have have : persona consistency : dialogue coherence Figure 1: Each response in a persona-based dialogue is mostly related to one persona sentence and its latest dialogue history utterance. Persona sentences in grey are redundant for all responses. a …

Content truncated to 500 words