A Survey of Controllable Text Generation using Transformer-based Pre-trained Language Models (2201.05337v5)

Abstract: Controllable Text Generation (CTG) is emerging area in the field of natural language generation (NLG). It is regarded as crucial for the development of advanced text generation technologies that better meet the specific constraints in practical applications. In recent years, methods using large-scale pre-trained LLMs (PLMs), in particular the widely used transformer-based PLMs, have become a new paradigm of NLG, allowing generation of more diverse and fluent text. However, due to the limited level of interpretability of deep neural networks, the controllability of these methods need to be guaranteed. To this end, controllable text generation using transformer-based PLMs has become a rapidly growing yet challenging new research hotspot. A diverse range of approaches have emerged in the recent 3-4 years, targeting different CTG tasks that require different types of controlled constraints. In this paper, we present a systematic critical review on the common tasks, main approaches, and evaluation methods in this area. Finally, we discuss the challenges that the field is facing, and put forward various promising future directions. To the best of our knowledge, this is the first survey paper to summarize the state-of-the-art CTG techniques from the perspective of Transformer-based PLMs. We hope it can help researchers and practitioners in the related fields to quickly track the academic and technological frontier, providing them with a landscape of the area and a roadmap for future research.

Controllable Text Generation with Transformer-Based Pre-trained LLMs: A Systematic Review

This survey paper provides a comprehensive overview of Controllable Text Generation (CTG) techniques leveraging Transformer-based Pre-trained LLMs (PLMs), marking an important stride in advancing Natural Language Generation (NLG). The review covers various state-of-the-art approaches for CTG, categorizing them based on the interaction level with PLMs, namely fine-tuning, retraining or refactoring, and post-processing. The primary aim is to bridge the gap in controllability observed in PLM-driven text generation, while emphasizing the delicate balance between maintaining high text quality and adhering to predefined constraints.

Key Approaches in Controllable Text Generation

1. Fine-Tuning:
   • This category encompasses techniques that adapt PLMs to meet specific control conditions with minimal resource overhead compared to training from scratch. Strategies like Adapted Modules and Prompt Learning are discussed, where model fine-tuning is pursued to guide the generative process towards specific attributes or styles. Reinforcement Learning (RL)-inspired and Instruction Tuning methods leverage human feedback or explicit instructions to enhance generation alignment with human intent.
2. Retraining/Refactoring:
   • Approaches in this category include either structural modifications of existing PLMs or training new models from scratch tailored for CTG. Techniques like CTRL and POINTER focus on integrating control tokens or using insertion-based generative workflows, respectively, to ensure compliance with lexical and syntactic constraints. While potent, such methods require substantial data and computational resources.
3. Post-Processing:
   • The third category emphasizes controlling text generation at decode time without altering the PLM. Strategies like Guided Methods and Trainable Models reposition the output probabilities of PLMs to emphasize desired characteristics in the text. By decoupling control modules from PLMs, these methods boast efficiency in training but often show increased inference costs and challenges in maintaining generation quality.

Evaluation Metrics

The survey highlights the dual nature of CTG evaluation: assessing both the alignment of generated text to controlled attributes and its linguistic quality. Established metrics such as BLEU for fluency and ROUGE for content overlap are complemented by CTG-specific measures like semantic consistency classifiers and human-centric evaluations for more subjective dimensions like relevance and coherence.

Challenges and Future Directions

The paper identifies several challenges in achieving robust CTG, including the balance between domain diversity and control specificity, the limitations of probabilistic modeling for long-text coherence, and the integration of PLMs with external knowledge repositories for better grounding in tasks needing world knowledge. Future directions suggested include exploring prompt-based learning, enhancing fine-grained decoding controls, and leveraging advanced linguistic and probabilistic models to improve text quality and adherence to constraints.

This survey serves not only as a comprehensive resource on the current CTG landscape but also as a blueprint for future research endeavors aimed at refining how transformer-based PLMs generate controlled, quality text. It establishes actionable insights into extending these frameworks towards more diverse and complex NLG applications, particularly in areas aligned with Artificial General Intelligence aspirations.