PCFG-Based Natural Language Interface Improves Generalization for Controlled Text Generation

Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 295–313 July 13-14, 2023 ©2023 Association for Computational Linguistics PCFG-Based Natural Language Interface Improves Generalization for Controlled Text Generation Jingyu Zhang Johns Hopkins University jzhan237@jhu.edu James Glass MIT glass@mit.edu Tianxing He University of Washington goosehe@cs.washington.edu Abstract Existing work on controlled text generation (CTG) assumes a control interface of categori- cal attributes. In this work, we propose a natu- ral language (NL) interface, where we craft a PCFG to embed the control attributes into natu- ral language commands, and propose variants of existing CTG models that take commands as input. In our experiments, we design tailored setups to test the model’s generalization abili- ties. We find our PCFG-based command gener- ation approach is effective for handling unseen commands compared to fix-set templates. Fur- ther, our proposed NL models can effectively generalize to unseen attributes (a new ability enabled by the NL interface), as well as unseen attribute combinations. Interestingly, in model comparisons, the simple conditional generation approach, enhanced with our proposed NL in- terface, is shown to be a strong baseline in those challenging settings. 1 Introduction With the advancement of large-scale pretraining, language models (LM) are now able to generate in- creasingly more realistic text (Radford et al., 2019; Brown et al., 2020; Rae et al., 2021; Hoffmann et al., 2022; Smith et al., 2022; Thoppilan et al., 2022). Therefore, how to control the generation of LMs has become an important research topic. In controlled text generation (CTG), a series of works (Keskar et al., 2019; Dathathri et al., 2020; Krause et al., 2021; Yang and Klein, 2021; Liu et al., 2021; Yu et al., 2021; Li et al., 2022, inter alia) propose model frameworks to generate text conditioned on some desired (user-specified) attribute a. These attributes, which depend on the datasets of interest, could be topic, formality, sentiment, etc. An important assumption behind this controlled generation setting is that the attributes are chosen from a fixed set (i.e., they are treated as categorical random variables). Although this setting is con- venient, it seriously limits the applications of the Figure 1: We explore generation models that take natu- ral language commands as input. For training, we use PCFG to embed categorical control attributes into natu- ral language commands. CTG system: (1) Since the attribute set is fixed dur- ing training, it would be impossible for the model to generalize to unseen options if used as-is. (2) This interface is not very human-friendly, because it could be difficult for users to navigate through the (possibly long) lists of options. Motivated by these limitations, in this work we propose a natu- ral language interface for CTG, illustrated in Fig- ure 1. With this change of interface, the input to the CTG model changes from one-hot vectors to natu- ral language commands (for short, commands). To efficiently train this system and enable it to general- ize, we design a probabilistic context-free grammar (PCFG) to embed categorical attributes into …

Content truncated to 500 words