Seen to Unseen: Exploring Compositional Generalization of Multi-Attribute Controllable Dialogue Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 14179–14196 July 9-14, 2023 ©2023 Association for Computational Linguistics Seen to Unseen: Exploring Compositional Generalization of Multi-Attribute Controllable Dialogue Generation Weihao Zeng1∗, Lulu Zhao1∗, Keqing He2, Ruotong Geng1 Jingang Wang2, Wei Wu2, Weiran Xu1∗ 1Beijing University of Posts and Telecommunications, Beijing, China 2Meituan, Beijing, China {zengwh,zhaoll,ruotonggeng,xuweiran}@bupt.edu.cn {hekeqing,wangjingang,wuwei}@meituan.com Abstract Existing controllable dialogue generation work focuses on the single-attribute control and lacks generalization capability to out-of-distribution multiple attribute combinations. In this paper, we explore the compositional generalization for multi-attribute controllable dialogue generation where a model can learn from seen attribute values and generalize to unseen combinations. We propose a prompt-based disentangled con- trollable dialogue generation model, DCG. It learns attribute concept composition by gener- ating attribute-oriented prompt vectors and uses a disentanglement loss to disentangle different attributes for better generalization. Besides, we design a unified reference-free evaluation framework for multiple attributes with different levels of granularities. Experiment results on two benchmarks prove the effectiveness of our method and the evaluation metric. 1 Introduction Recently, large pre-trained language models (PLMs) like DialoGPT (Zhang et al., 2020), BlenderBot (Roller et al., 2020) and Meena (Adi- wardana et al., 2020) can produce fluent and rel- evant responses for dialogue contexts. However, the generated responses are often uninformative and factual inconsistent. Hence, controllable dia- logue generation (CDG) is proposed to guide di- alogue generation towards the desired attributes such as emotions (Zhou et al., 2018), acts (Li et al., 2017), and personas (Zhang et al., 2018). Previous work focused on directly fine-tuning the large-scale PLMs (Keskar et al., 2019) or using an extra at- tribute discriminator (Krause et al., 2021; Dathathri et al., 2019) to guide generation. The former is ex- pensive and requires extensive annotated attribute labels. The decoding of the latter is computation- ally intensive, reducing the response fluency and generation speed. ∗The first two authors contribute equally. Weiran Xu is the corresponding author. 1 2 3 4 5 6 7 8 0.0 0.2 0.4 0.6 0.8 unseen multi-attribute combination seen multi-attribute combination (a) E-ACC 1 2 3 4 5 6 7 8 0.0 0.2 0.4 0.6 0.8 unseen multi-attribute combination seen multi-attribute combination (b) A-ACC Figure 1: The difference of controllability scores on seen and unseen multi-attribute combinations of CTRL (Keskar et al., 2019). E-ACC and A-ACC denote emo- tion accuracy and act accuracy. Although these methods have made some progress in CDG, most of them focus on single- attribute generation where there is only one at- tribute label like happiness in emotion and pay less attention to the multi-attribute generation, which is a more practical setting. Therefore, we are commit- ted to filling this gap in CDG. Noted that different from single-attribute, the control signal of the multi- attribute generation is a combination of multiple values from different attributes, which faces the challenge of lacking sufficient annotated attribute- specific data. We also find state-of-the-art meth- ods for multi-attribute controllable text generation (Yang et al., 2022; …

Content truncated to 500 words