Controllable Text Generation via Probability Density Estimation in the Latent Space

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12590–12616 July 9-14, 2023 ©2023 Association for Computational Linguistics Controllable Text Generation via Probability Density Estimation in the Latent Space Yuxuan Gu†, Xiaocheng Feng†‡, Sicheng Ma†, Lingyuan Zhang†, Heng Gong†, Weihong Zhong†, Bing Qin†‡ †Harbin Institute of Technology ‡ Peng Cheng Laboratory {yxgu,xcfeng,scma,lyzhang,hgong,whzhong,qinb}@ir.hit.edu.cn Abstract Previous work on controllable text generation has explored the idea of control from the latent space, such as optimizing a representation with attribute-specific classifiers or sampling one from relevant discrete samples. However, they cannot effectively model a complex space with diverse attributes, high dimensionality, and asymmetric structure, leaving subsequent con- trols unsatisfying. In this work, we propose a novel control framework using probability den- sity estimation in the latent space. Our method utilizes an invertible transformation function, the Normalizing Flow, that maps the complex distributions in the latent space to simple Gaus- sian distributions in the prior space. Thus, we can perform sophisticated and flexible controls in the prior space and feed the control effects back into the latent space owing to the bijection property of invertible transformations. Exper- iments on single-attribute and multi-attribute control reveal that our method outperforms sev- eral strong baselines on attribute relevance and text quality, achieving a new SOTA. Further analysis of control strength adjustment demon- strates the flexibility of our control strategy1. 1 Introduction Controllable text generation, a fundamental issue in natural language generation, refers to generat- ing fluent and attractive sentences conditioned on target attributes (Zhang et al., 2022a). With the development of pre-trained language models (Zhao et al., 2023), early work explores converting gen- erative language models to conditional models by altering their parameters via fine-tuning (Ziegler et al., 2019; Keskar et al., 2019) or reinforcement learning (Khalifa et al., 2020). Due to the high cost of modifying parameters (Brown et al., 2020; Zhang et al., 2022b), current control approaches prefer leaving pre-trained language models fixed (Dathathri et al., 2020; Krause et al., 2021). 1 https://github.com/HappyGu0524/MultiControl. Figure 1: Illustration of methods controlling in Latent Space. Orange background denotes the latent space. Blue and red represent two attributes. Prefix-Tuning represents attributes with points in latent space and com- poses them by interpolation. LatentOps uses classifiers to estimate continuous distributions of attributes and control by optimizing in latent space. Discrete maps sentences to discrete samples in latent space and con- trols with direct searching. Our method deploys proba- bility density estimation by transforming the complex latent space into a well-formed prior space, where com- mon control strategies can be more effective. See §A for more details of the latent space’s defects. Recent studies perform impressive control by influencing the fixed language model from the la- tent space (Yu et al., 2021; Qian et al., 2022) with prefix-tuning (Li and Liang, 2021). However, in- efficient and unreliable modeling of the complex latent space remains a problem that plagues control performance. As shown in the left part of Figure 1, Gu et al. (2022b) reveal …

Content truncated to 500 words