An Extensible Plug-and-Play Method for Multi-Aspect Controllable Text Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 15233–15256 July 9-14, 2023 ©2023 Association for Computational Linguistics An Extensible Plug-and-Play Method for Multi-Aspect Controllable Text Generation Xuancheng Huang1† Zijun Liu2,4† Peng Li3,5 Tao Li1 Maosong Sun2,4∗ Yang Liu2,3,4,5∗ 1Meituan, China 2Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University, Beijing, China 3Institute for AI Industry Research (AIR), Tsinghua University, Beijing, China 4Beijing National Research Center for Information Science and Technology 5Shanghai Artificial Intelligence Laboratory, Shanghai, China Abstract Recently, multi-aspect controllable text genera- tion that controls the generated text in multiple aspects (e.g., sentiment, topic, and keywords) has attracted increasing attention. Although methods based on parameter efficient tuning like prefix-tuning could achieve multi-aspect controlling in a plug-and-play way, the mutual interference of multiple prefixes leads to sig- nificant degeneration of constraints and limits their extensibility to training-time unseen as- pect combinations. In this work, we provide a theoretical lower bound for the interference and empirically found that the interference grows with the number of layers where prefixes are inserted. Based on these analyses, we propose using trainable gates to normalize the interven- tion of prefixes to restrain the growing inter- ference. As a result, controlling training-time unseen combinations of aspects can be realized by simply concatenating corresponding plugins such that new constraints can be extended at a lower cost. In addition, we propose a unified way to process both categorical and free-form constraints. Experiments on text generation and machine translation demonstrate the superior- ity of our approach over baselines on constraint accuracy, text quality, and extensibility.1 1 Introduction Multi-aspect controllable text generation (MCTG), which aims at generating fluent text while satis- fying multiple aspects of constraints simultane- ously, has attracted increasing attention in recent years (Chan et al., 2021; Qian et al., 2022; Gu et al., 2022). To effectively control diverse aspects such as sentiment, topic, and detoxification, extensive ef- forts have been devoted to the task, including meth- ods based on conditional generative model (Keskar † indicates equal contribution. * Corresponding authors: M.Sun (sms@tsinghua.edu .cn) and Y.Liu (liuyang2011@tsinghua.edu.cn) 1The source code is available at https://github.com/ THUNLP-MT/PromptGating4MCTG et al., 2019), decoding-time regulation (Lin and Riedl, 2021; Kumar et al., 2021), and parameter efficient tuning (Qian et al., 2022; Gu et al., 2022). Despite their effectiveness, existing methods still suffer from low extensibility. Ideally, suppose a multi-aspect controllable text generation system has learned how to control sentiment, topic and keywords separately, it should be extensible to any combinations of the three aspects, e.g., generating a sports-themed sentence with negative sentiment containing keywords “New York” (see Figure 1). Moreover, an extensible system should also be eas- ily extended to control new aspects in a plug-and- play way. However, it is non-trivial for existing methods to achieve this goal. Specifically, the dedi- cated conditional generative models (Keskar et al., 2019) mostly need to be trained from scratch or finetuned when facing new aspect combinations. The decoding-time regulation based methods (Lin and Riedl, 2021; …

Content truncated to 500 words