TL;DR
This paper surveys the rapidly evolving field of sentiment-controlled text generation, proposing a novel categorization scheme for existing techniques. It provides a comprehensive overview and comparison of state-of-the-art methods, highlighting their strengths, weaknesses, and future research directions.
What Problem Does This Paper Address?
Large language models (LLMs) like GPT and BERT have revolutionized text generation, but they often struggle with controlling specific attributes of the generated text, leading to outputs that can be irrelevant, incorrect, or inappropriate. This paper focuses on a crucial aspect of this control problem: sentiment. While LLMs can generate fluent text, ensuring it consistently conveys a desired sentiment (positive, negative, neutral, etc.) remains a significant challenge. Existing research has explored various methods to achieve this, but a clear, organized understanding of these techniques and their comparative performance is lacking. This survey aims to fill that gap by systematically categorizing and analyzing the state-of-the-art in sentiment-controlled text generation.
Methods
The authors propose a two-part categorization scheme for sentiment-controlled text generation techniques:
- Control Attribute Types: They identify three main categories of attributes that can be controlled:
- Polarity Control: Directly controlling sentiment (positive/negative) or related concepts like toxicity.
- Content Control: Ensuring the presence of specific words, topics, or characters.
- Syntactic Control: Manipulating sentence structure, tense, or dialogue acts.
- Control Implementation Techniques: They categorize how control is achieved:
- Complete Training: Training models from scratch for the specific task.
- Model Fine-Tuning: Adapting pre-trained language models (PLMs) to the control task.
- Disentanglement: Extracting attribute representations to guide generation.
- Modification of Token Distribution: Adjusting the probability of generating certain tokens.
- Hybrid: Combinations of the above techniques.
The paper systematically searched the ACL Anthology for relevant papers, filtering for those presenting new methods for sentiment control. They then categorized these selected papers (19 in total) using their proposed scheme and analyzed them based on performance, efficiency, and generality.
Datasets
The paper mentions several datasets used for training and evaluation, including:
- ZHIHU: Used in some Complete Training methods.
- Movie Reviews (Socher et al., 2013): Used for evaluating attribute relevance.
- Customer Reviews (Hu and Liu, 2004): Used in some Complete Training methods.
- Story Commonsense: Used in some Complete Training methods.
- PPLM Prompts: A set of prompts used for evaluating perplexity and attribute relevance in several methods.
- Yelp: Used for fine-tuning classifiers in some methods.
- OpenWebText Corpus: Used for evaluating techniques.
- ROCStory (Rashkin et al., 2018): Used in some Hybrid methods.
- weibo.com (Shang et al., 2015): Used in some Hybrid methods.
- Bookcorpus (Zhu et al., 2015): Used in some Hybrid methods.
- IMDb: Used for fine-tuning classifiers.
- SenTube (Uryupina et al., 2014): Used in some Hybrid methods.
- SST-2: Used for sentiment classification.
The paper does not introduce new datasets but rather surveys existing work that utilizes these common benchmarks.
Key Insights and Ideas
- Categorization is Key: The proposed categorization scheme for control attributes and implementation techniques provides a much-needed framework for understanding and comparing diverse sentiment control methods.
- Rapid Growth: The number of papers on sentiment-controlled text generation has significantly increased from 2019 to 2022, indicating a growing research interest.
- Dominance of Fine-Tuning and Hybrid: Model Fine-Tuning and Hybrid techniques appear to be popular approaches, likely due to their ability to leverage powerful pre-trained models.
- Trade-offs Exist: While many methods aim for multi-attribute control, there's often a trade-off between control granularity, fluency, and diversity.
- Evaluation Challenges: A lack of a standardized evaluation procedure and consistent use of metrics across papers makes direct comparison difficult. Different classifiers for attribute relevance and varying language models for perplexity evaluation are noted as significant issues.
- Limited Attribute Combinations: While sentiment is often controlled, its combination with other attribute types (like content or syntactic control) is less explored than might be expected.
Key Results
The paper presents comparative tables (Tables 2, 3, and 4) summarizing the performance, generality, and parameter efficiency of the surveyed techniques.
- Performance:
- Yu et al. (2021) showed good variety in generated texts while maintaining fluency when evaluated using PPLM prompts.
- Landsman et al. (2022) achieved the highest attribute relevance and diversity on OpenWebText prompts for both positive and negative sentiment targets.
- Generality & Efficiency:
- Some techniques are highly specialized and require significant modifications for new attributes or tasks.
- Model-agnostic techniques (like those by Landsman et al. and Dathathri et al.) are advantageous as they can be applied to any PLM.
- Modification of Token Distribution techniques generally require fewer trainable parameters, with some methods (Gu et al., 2022a; Landsman et al., 2022) achieving zero trainable parameters by using sampling procedures.
Limitations and Future Work
The authors identify several limitations and suggest future directions:
- Standardized Evaluation: A pressing need for a unified evaluation procedure and standardized metrics to enable fair comparisons across different methods and datasets.
- Broader Attribute Combinations: Exploring combinations of sentiment with other control attributes, particularly content control attributes beyond topic and toxicity, and syntactic control.
- Model Generalization: Extending models to generalize across different domains and tasks (e.g., poetry generation, machine translation).
- Fine-grained Control: Developing methods for more nuanced control, potentially by exploring correlations between different attributes or extending control to scenarios like table-to-text generation.
- Exploring New Tasks: Applying sentiment control techniques to novel tasks such as poetry generation, machine translation, and intelligent education agents.
Summary updated