Fine-Grained Controllable Text Generation Using Non-Residual Prompting

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 6837 - 6857 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Fine-Grained Controllable Text Generation Using Non-Residual Prompting Fredrik Carlsson∗ Joey ¨Ohman† Fangyu Liu‡ Severine Verlinden† Joakim Nivre∗ Magnus Sahlgren† ∗Research Institutes of Sweden fredrik.carlsson@ri.se joakim.nivre@ri.se †AI Sweden joey.ohman@ai.se severine.verlinden@ai.se magnus.sahlgren@ai.se ‡University of Cambridge fl339@cam.ac.uk Abstract The introduction of immensely large causal language models (CLMs) has rejuvenated the interest in open-ended text generation. How- ever, controlling the generative process for these Transformer-based models is at large an unsolved problem. Earlier work has ex- plored either plug-and-play decoding strate- gies or more powerful but blunt approaches such as prompting. There hence currently ex- ists a trade-off between fine-grained control and the capability for more expressive high- level instructions. To alleviate this trade-off, we propose an encoder-decoder architecture that enables intermediate text prompts at ar- bitrary time steps. We propose a resource- efficient method for converting a pre-trained CLM into this architecture and demonstrate its potential in various experiments, including the novel task of contextualized word inclusion. Our method provides strong results in multi- ple experimental settings, proving itself to be both expressive and versatile.1 1 Introduction A causal language model (CLM) is a language model trained using a simple next-token predic- tion objective. Current CLMs are typically based on the Transformer architecture (Vaswani et al., 2017), which has resulted in unprecedented text generation capabilities (Radford et al., 2018a,b; Brown et al., 2020). Even so, the generation pro- cess of a CLM is difficult to control, as one is forced to gradually decode the next-step predic- tion one token at a time. This inhibits the applica- bility of CLMs when one intends for the generated text to fulfill certain criteria, and not only be a lin- guistically sound continuation in a given context. Being able to control the text generation process is crucial for many real-world applications. As a straightforward example, we may want to control 1Code and models: https://Github.com/ FreddeFrallan/Non-Residual-Prompting the generated text to counter the many biases that modern CLMs have been shown to possess (Bor- dia and Bowman, 2019). However, most applica- tions require a greater degree of control, as one often wishes to steer the text generation in a spe- cific direction, such as generating a story to a given plot (Li et al., 2013; Yao et al., 2019; Riedl, 2021), or sticking to a certain topic (Keskar et al., 2019). Some areas require stringent and fine-grained con- trol, as the many data-to-text tasks (Gardent et al., 2017; Lepp¨anen et al., 2017; Koncel-Kedziorski et al., 2019), which necessitates that the generated text mediates very specific information and facts. Due to this apparent need for controllable text generation, recent work (see Section 2.1) has ex- plored different methods to steer and constrain the generation process of a CLM. There are mainly two lines of research in this area. The more tra- ditional approach focuses on fine-grained control and how to steer the …

Content truncated to 500 words