Source-linked AI summary
A Survey of Controllable Text Generation using Transformer-based Pre-trained Language Models
Hanqing Zhang, Haolin Song, Shaoyu Li, Ming Zhou, Dawei Song
TL;DR
Practical NLG systems need controllability, while Transformer-based PLMs introduce powerful but challenging foundations for CTG. This paper systematically reviews CTG tasks, PLM-based approaches, and evaluation methods, then identifies challenges and future directions. It concludes that PLMs create significant opportunities, although decoding-time control, evaluation, and dedicated benchmarks remain constrained.
Problem
CTG must satisfy task-specific constraints, but the field lacked a comprehensive review of PLM-based tasks, approaches, evaluation methodologies, and current challenges.
Method
The paper provides a systematic critical review of CTG applications, evaluation methods, Transformer-based PLM approaches, challenges, and future directions.
Results
The survey synthesizes the field's landscape and identifies key challenges and promising future directions for PLM-based CTG.
Takeaways & Limitations
Large-scale PLMs provide unprecedented opportunities for CTG research and motivate further work toward the field's development.
Takeaways & Limitations
Decoding-time methods often have coarse guidance, reduced generated-text quality, and limited applicability to fine-grained or multi-attribute control.
Abstract
from arXiv · showhide
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 language models (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.
1 INTRODUCTION
Controllable text generation is a fundamental NLG issue because applications require outputs to satisfy task-specific constraints. Transformer-based PLMs have become central to CTG research, motivating this survey's comprehensive review of tasks, approaches, evaluation, challenges, and future directions.
- CTG requires generated text to satisfy application- and user-imposed constraints, including storylines, dialogue attributes, balanced data distributions, and avoidance of offensive content.These requirements make controllability important for practical NLG applications.
- Deep-learning methods enabled end-to-end, data-driven CTG, following earlier sequential and style-embedding approaches and later generative-model methods.The surveyed examples include VAEs, GANs, and energy-based models.
- Large-scale Transformer-based PLMs became a new NLP paradigm by learning semantic and syntactic knowledge from large corpora and requiring fine-tuning for downstream tasks.The passage identifies BERT, RoBERTa, GPT, T5, and mBART as examples.
- Existing CTG surveys lacked systematic coverage of representative tasks, main approaches, evaluation methodologies, and the latest large-scale PLM-based methods.This gap motivates a comprehensive critical review and roadmap.
- The paper reviews CTG tasks, evaluation metrics, PLM-based approaches, field challenges, and promising future research directions.Its stated goal is to help researchers and practitioners capture the field's overall picture and cutting-edge methods.
2 AN INTRODUCTION TO CONTROLLABLE TEXT GENERATION AND PRE-TRAINED LANGUAGE MODELS
CTG generates text under explicit controlled conditions while preserving general language qualities such as fluency, rationality, and readability. The survey organizes control conditions, task categories, and Transformer-based PLM types to frame the field's applications and methods.
- 2.1 Controllable Text Generation: A typical CTG system maps a controlled condition and source text through a generative model to output text satisfying the control condition.The IPO formulation defines input, process, and output components.
- 2.1 Controllable Text Generation: Control conditions are categorized as semantic, structural, or lexical constraints.Examples include sentiment or topic, text structure, and required keywords.
- 2.2 Tasks and Applications Involving CTG: CTG tasks span attribute-based generation, dialogue, storytelling, data augmentation, debiasing, and other application categories with distinct controlled aspects.The survey's overview enumerates seven task categories and describes their inputs, outputs, and representative references.
- 2.3 Transformer-based Pre-trained Language Models: PLMs use large-scale unlabeled data and achieve strong downstream performance after fine-tuning on relatively small amounts of supervised data.Transformer-based PLMs are the focus of this survey.
- 2.3 Transformer-based Pre-trained Language Models: Auto-encoding and sequence-to-sequence PLMs use bidirectional attention, whereas auto-regressive models rely on causal attention.The survey discusses BERT, RoBERTa, and ERNIE as AE examples; GPT as AR; and T5 and mBART as Seq2Seq examples.
3 MAIN APPROACHES TO PLM-BASED CTG
PLM-based CTG methods address the challenge of controlling powerful but relatively opaque language models by injecting control signals through distinct strategies. The survey organizes these approaches into three broad categories and reviews their mechanisms, benefits, and limitations.
- Overview: PLM-based CTG methods are broadly divided into fine-tuning, retraining or refactoring, and post-processing according to how control signals interact with the model.The overview further divides these categories into subclasses.
- Fine-tuning: Inverse Prompt improves prompt–generation relevance by reverse-predicting prompts for candidate tokens, but its repeated computation can make generation impractical.The survey reports up to around 10 minutes to generate a seven-word rhyming poem in tests using the provided source code.
- Fine-tuning: Fine-tuning methods adapt part or all of a PLM’s parameters to generate text satisfying specified control conditions.Instruction tuning and attribute-guided methods exemplify this strategy.
- Post-Processing: Post-processing methods keep PLM parameters fixed and rerank token distributions or sample candidates to steer generation toward desired attributes.Examples include discriminator or energy-based guidance and methods such as Mix and Match.
- Post-Processing: Guided and trainable post-processing strategies are flexible and theoretically grounded, but can reduce efficiency or text quality because they require extra computation or iterative processing.Reported limitations include excessive inference time, high computational cost, and early-stage text-quality issues.
4 EVALUATION METHODS
CTG evaluation combines general NLG measures of text quality with task-specific measures of controlled-element satisfaction. The survey organizes general evaluation into human-centric and automatic approaches, highlighting trade-offs among precision, cost, speed, and consistency.
- Evaluation overview: CTG models require both general NLG metrics and task-specific metrics measuring satisfaction of controlled elements.General metrics assess generated-text quality, while CTG-specific metrics assess consistency with the imposed controls.
- General NLG Evaluation Metrics: General text quality is assessed through fluency, factuality, grammar, and diversity, using human or machine assessors.These assessment modes are categorized in Figure 7.
- Human-centric evaluation: Human-centric evaluation includes direct judgments such as binary ratings, Likert scales, and rankings, plus indirect downstream-task evaluations.Indirect evaluation measures user task success or system purpose success rather than judging text quality directly.
- Automatic evaluation: Automatic NLG metrics compare generated text with references using lexical-, syntactic-, or semantic-based similarity measures.Semantic metrics such as BERTScore and BLEURT use PLM representations or training to capture meaning beyond surface overlap.
- Automatic evaluation: BLEU-n emphasizes precision through generated-text n-grams, whereas ROUGE-n emphasizes recall through reference-text n-grams.The denominator distinguishes the two metrics, and larger ROUGE-n indicates better recall-oriented quality.
- Evaluation trade-offs: Human-centric evaluation is more precise but costly, time-consuming, difficult to quality-control, and potentially inconsistent; automatic evaluation is faster and cheaper but less precise.Semi-automatic methods combine both approaches but still require substantial human judgments or labeling.
5 CHALLENGES AND FUTURE DIRECTIONS
The survey identifies persistent challenges in controllable text generation, including domain adaptation, fine-grained decoding, long-range coherence, knowledge grounding, evaluation, and dedicated benchmarks. It proposes future directions spanning prompting, finer-grained control, classical generative theory and linguistic knowledge, external knowledge, improved metrics, and new CTG tasks.
- Challenges: PLM-based models struggle to preserve domain diversity while satisfying control conditions, especially with few or zero domain-specific samples.The survey relates this difficulty to catastrophic forgetting during adaptation.
- Challenges: Decoding-time control is inexpensive and preserves pretrained-model characteristics, but decoupled discriminators can produce coarse guidance and reduced text quality.These methods are also difficult to apply directly to fine-grained data-to-text and multi-attribute control.
- Challenges: Locally normalized autoregressive PLMs have difficulty maintaining semantic and controlled-condition coherence across paragraphs and documents.The survey calls for global normalization that supports both local and global control.
- Challenges: The field lacks an objective, accurate, and comprehensive evaluation mechanism compatible with human judgment for both text quality and control satisfaction.The survey treats reliable evaluation as a bottleneck for advancing controllable generation.
- Future Directions: Prompt-based learning may support few-shot or zero-shot adaptation, broader text fields, and reduced catastrophic forgetting in controlled generation.The proposed future directions also include co-training guided and generative models, unified multi-attribute control, and integration with classical generative theory and linguistic knowledge.
- Future Directions: Future work should incorporate retrieved information or knowledge graphs, develop metrics for fluency, diversity, coherence, and fidelity, and create dedicated CTG benchmarks with diverse control requirements.The survey also proposes AGI-oriented benchmarks and tasks for accurate and reliable language-model outputs.
6 CONCLUSIONS
The paper comprehensively surveys applications, approaches, and evaluation methodologies for controllable text generation based on large-scale pretrained language models. It identifies key challenges and future directions, aiming to provide a field landscape and roadmap for researchers and practitioners.
- Conclusions: The survey covers typical applications, main approaches, and evaluation methodologies for controllable text generation based on large-scale pretrained language models.It presents these components alongside a critical analysis of existing methods.
- Conclusions: The authors identify key challenges and promising future directions for controllable text generation research.They frame large-scale pretrained language models as opportunities for further development of the field.
- Conclusions: The survey aims to provide researchers and practitioners with a clear picture of the field and a roadmap for moving forward.