Source-linked AI summary
Deep Learning for Text Style Transfer: A Survey
Di Jin, Zhijing Jin, Zhiting Hu, Olga Vechtomova, Rada Mihalcea
TL;DR
Text style transfer seeks to control textual attributes while preserving content, but its literature uses diverse datasets, methods, and evaluation practices. This paper surveys more than 100 neural TST works, organizing task formulations, datasets, evaluation, parallel and non-parallel methods, applications, and research directions. The survey provides a structured reference and standardization perspectives while identifying important methodological, evaluation, computational, and scope challenges.
Problem
TST research has grown across diverse benchmark datasets, methodological frameworks, and evaluation metrics, creating a need for synthesis and potential standardization.
Method
The paper systematically reviews more than 100 neural TST works and categorizes their formulations, datasets, evaluation, methods, applications, and research agenda.
Results
The survey distills existing literature into methodological frameworks for parallel and non-parallel data and provides standard-practice discussions for task settings, datasets, and evaluation.
Takeaways & Limitations
The survey serves as a reference for TST researchers and connects the task to downstream applications, other NLP tasks, and ethical considerations.
Takeaways & Limitations
Pseudo-parallel corpus construction can require sufficient matching samples, successful bootstrapping, convergence, and costly iterative retraining.
Abstract
from arXiv · showhide
Text style transfer is an important task in natural language generation, which aims to control certain attributes in the generated text, such as politeness, emotion, humor, and many others. It has a long history in the field of natural language processing, and recently has re-gained significant attention thanks to the promising performance brought by deep neural models. In this paper, we present a systematic survey of the research on neural text style transfer, spanning over 100 representative articles since the first neural text style transfer work in 2017. We discuss the task formulation, existing datasets and subtasks, evaluation, as well as the rich methodologies in the presence of parallel and non-parallel data. We also provide discussions on a variety of important topics regarding the future development of this task. Our curated paper list is at https://github.com/zhijing-jin/Text_Style_Transfer_Survey
1 Introduction
Text style transfer controls stylistic attributes while preserving content, supporting more user-centered and personalized language generation. This survey organizes a diverse and rapidly growing neural literature across task settings, datasets, methods, evaluation, and future directions.
- Text style transfer controls text attributes while preserving content, with applications including persona-consistent intelligent bots and writing assistance.Attributes can include formality, politeness, simplicity, personality, emotion, and genre.
- The formal task models p(x′|a, x), generating a target-style sentence from a source sentence and source attribute while conditioning on the target attribute.Unlike style-conditioned language modeling, TST uses the source sentence to constrain content.
- The survey addresses the diversity of benchmark datasets, methodological frameworks, and evaluation metrics in the growing TST literature.It aims to summarize and potentially standardize terminology, problem definitions, datasets, and metrics.
- The review covers more than 100 deep learning-based TST papers and categorizes methods for both parallel and non-parallel data.It also presents unified methodological frameworks for these settings.
- The survey proposes a research agenda spanning broader styles, improved methods and evaluation, weaker dataset assumptions, downstream applications, and ethical impacts.The envisioned applications connect TST with other NLP tasks and specialized uses.
2 What Is Text Style Transfer?
The survey defines TST as controlling text attributes while preserving content, while distinguishing linguistic and data-driven notions of style. It then scopes common subtasks and datasets used by neural approaches.
- How to Define Style?: Linguistic style concerns how semantics is expressed through choices such as words, syntax, metaphors, and discourse structure.The definition includes personal utterance characteristics and stylistic organization beyond the sentence level.
- How to Define Style?: Data-driven style treats attributes that vary across corpora as style and invariants as content, allowing sentiment and topic preferences within scope.This broader definition reflects how many neural TST datasets are constructed.
- Task Formulation: Most surveyed methods require style-specific corpora for each attribute, either parallel or non-parallel, making corpus availability a core assumption.The survey notes that future work may remove this assumption.
- Existing Subtasks with Datasets: Common subtasks span stylistic attributes such as formality and politeness and content-related attributes such as sentiment and topics.The survey organizes these subtasks alongside their corresponding datasets and attribute values.
- Existing Subtasks with Datasets: The politeness dataset contains 1.39 million automatically labeled Enron instances and mainly represents North American English because politeness is culture-dependent.Its cultural scope limits direct generalization to other varieties of English or cultures.
3 How to Evaluate Style Transfer?
Style-transfer evaluation commonly measures transferred style strength, semantic preservation, and fluency using automatic and human methods. The survey highlights important limitations, including metric blind spots, weak correlations, reference scarcity, and the cost and irreproducibility of human judgments.
- Evaluation commonly considers transferred style strength, semantic preservation, and fluency.
- Automatic Evaluation: Automatic evaluation is economical, reproducible, and scalable, but each metric addresses only certain aspects and has intrinsic blind spots.
- Automatic Evaluation: BLEU can reward copying the input, correlate weakly with human judgments, and is unavailable when datasets lack human-written references.
- Automatic Evaluation: Style-classifier accuracy may diverge from human judgments: it correlates on Yelp and Captions but has almost no correlation on Amazon.
- Automatic Evaluation: Perplexity has disputed correlation with human fluency ratings and is sensitive to sentence length, word frequency, domain shift, and model configuration.
- Human Evaluation: Human evaluation can provide overall or criterion-specific judgments, but it is time-consuming, costly, subjective, and difficult to compare across studies.
4 Methods on Parallel Data
Parallel-data methods predominantly use neural sequence-to-sequence models, with extensions that add auxiliary objectives, inference controls, data augmentation, or pretrained language models. GPT-2 use is reported to improve BLEU by at least 7 points with human references.
- Parallel-data TST methods are distinguished from non-parallel methods by whether corresponding sentences with different styles are available.
- Most parallel-data methods adopt encoder-decoder neural sequence-to-sequence architectures.
- Multi-Tasking: Multi-tasking adds classifier-guided, self-reconstruction, and cycle losses to encourage target-style conformity, input reconstruction, and transfer consistency.
- At least 7 BLEU scores of improvement with human references are reported when using pretrained GPT-2.
- Inference Techniques: Negative lexically constrained decoding targets source words requiring replacement and can be applied without retraining the style-transfer model.
- Data Augmentation: Back-translation and scraped online text are used to augment scarce parallel data with pseudo-parallel training examples.
5 Methods on Non-Parallel Data
Because parallel style-transfer data is difficult or impossible to obtain for many styles, most methods use non-parallel mono-style corpora. The survey groups these methods into disentanglement, prototype editing, and pseudo-parallel corpus construction.
- Parallel data is difficult to obtain and can be impossible to crowdsource for some style pairs.
- Most non-parallel TST methods assume mono-style corpora and investigate deep learning under this data constraint.
- The survey organizes non-parallel methods into disentanglement, prototype editing, and pseudo-parallel corpus construction.
5.1 Disentanglement
Disentanglement-based TST models encode text into latent representations, manipulate those representations to change attributes, and decode the result. The survey organizes these models by encoder-decoder backbone, latent manipulation strategy, and style/content losses.
- Disentanglement-based models encode x into z, manipulate z into z′, and decode z′ into x′ with the target attribute.
- The survey’s workflow selects an encoder-decoder backbone, a latent-representation manipulation method, and appropriate loss functions.
- Encoder-Decoder Training Method: The main backbone choices are auto-encoders, variational auto-encoders, and generative adversarial networks.
- Encoder-Decoder Training Method: Denoising auto-encoding corrupts inputs by dropping, shuffling, or masking words before reconstruction, helping avoid blind copying.Pretraining has increased DAE’s popularity because it can improve semantic preservation and fluency.
- Disentanglement Methods: Latent manipulation methods include latent representation editing, attribute code control, and latent representation splitting.
- Disentanglement Methods: Attribute code control separates attribute-independent information in z from attribute value a, then uses a structured code to control generation.The code can act as the decoder’s initial state or embedding.
- Disentanglement Objectives: Disentanglement losses aim to make the target attribute exclusively controlled by a while attribute-independent information is exclusively captured by z.Style-oriented losses support the first aim, whereas content-oriented losses support the second.
- Style-Oriented Losses: Attribute Classifier on Outputs trains generated sentences to carry the target attribute according to a pretrained attribute classifier.It can use Gumbel-softmax training or reinforcement learning with policy gradients.
5.2 Prototype Editing
Prototype editing performs text style transfer through an explicit delete–retrieve–generate pipeline. It offers controllability and interpretability but depends on identifying attribute markers and matching suitable target markers.
- The approach connects neural prototype editing with earlier word-replacement methods using dictionaries, WordNet, rules, hypernyms, and definitions.
- Prototype editing explicitly transfers style by detecting and deleting source markers, retrieving target markers, and infilling a fluent sentence.
- Attribute Marker Detection: Attribute-marker extraction is non-trivial and uses frequency-ratio, attention-based, or fusion methods.
- Attribute Marker Detection: Frequency-ratio methods identify markers from n-gram statistics comparing their association with alternative attributes.Examples use relative co-occurrence frequency or ratios of mean TF-IDF.
- Attribute Marker Detection: Attention-based methods treat words with above-average attention from an attribute classifier as style markers.The cited classifier architectures include LSTM and BERT, with BERT reported to achieve higher marker-detection accuracy.
- Attribute Marker Detection: Fusion methods prioritize frequency-ratio markers and use attention-based detection as a backup or to reduce false positives.
- Attribute Marker Detection: Word-importance scoring addresses limitations from imperfect classifiers and unclear relationships between attribute and attention scores.It measures a token’s importance from the change in attribute probability after deleting that token.
- Target Attribute Retriever: Target markers are commonly retrieved by matching the source content template with the most similar template in the opposite-attribute corpus.Previous work often uses cosine similarity between sentence embeddings for template matching.
5.3 Pseudo-Parallel Corpus Construction
Pseudo-parallel corpus construction supplies aligned training signals for TST when only mono-style corpora are available. Retrieval matches semantically similar sentences, while generation-based methods iteratively create and reuse translated pairs.
- Pseudo-parallel data provide additional training signals for TST and are constructed mainly through retrieval-based or generation-based methods.
- Retrieval-Based Corpora Construction: Retrieval-based construction extracts aligned pairs from two mono-style corpora by matching semantically similar sentences.The initial pseudo corpora can be formed using cosine similarity of pretrained sentence embeddings.
- Generation-Based Corpora Construction: Generation-based construction uses iterative back-translation to generate pseudo-parallel corpora.
- Generation-Based Corpora Construction: IBT initializes two opposite-direction transfer models, generates pseudo-pairs in each direction, and retrains both models on those generated pairs.
- Generation-Based Corpora Construction: Random initialization is simple but subject to randomness and may fail to bootstrap well.An alternative initializes the process using an unsupervised word-to-word translation table.
- Generation-Based Corpora Construction: Iterative back-translation can diverge because no constraint ensures that each iteration produces better pseudo-parallel corpora.Additional attribute-classification losses or a checking mechanism using Word Mover Distance are described as ways to enhance convergence.
6 Research Agenda
The survey proposes a research agenda spanning broader style settings, stronger methodologies, fewer style-specific data assumptions, and more interpretable evaluation. It also identifies concrete limitations of current non-parallel approaches and opportunities to combine controllable traditional NLG structures with deep learning.
- 6.1 Expanding the Scope of Styles: Future TST research should expand beyond common styles toward aspect-based, authorship, document-level, and domain-adaptive transfer.The survey also mentions factual-to-empathetic, non-native-to-native, and sentence-disambiguation transfers as promising directions.
- 6.1 Expanding the Scope of Styles: A key open challenge is handling cases where style and semantics are interwoven rather than approximately separable.The current problem setting simplifies transfer by assuming that attribute and semantics can be separated.
- 6.2 Methodology: Non-parallel methods face distinct constraints: disentanglement needs inductive biases or supervision, prototype editing depends on markers and lexical overlap, and pseudo-parallel construction may be costly or fail to converge.Disentanglement also becomes more computationally expensive as the number of target styles increases.
- 6.2 Methodology: Prototype editing can achieve relatively high BLEU partly by copying input text, and it is especially suited to tasks with identifiable attribute markers and shared templates.Its retrieval and generation stages become problematic for confounded content-style data, low-overlap rewrites, or many styles.
- 6.2 Methodology: Combining traditional NLG planning with neural models offers a controllable, interpretable pipeline that can preserve templates, retrieve attributes, and better ensure information truthfulness.The survey presents this combination as easier to tune, debug, and improve than a purely black-box approach.
- 6.3 Loosening Data Assumptions: Future work could loosen style-specific corpus requirements through prompts, style-specific templates, synthetic data, or unsupervised discovery of styles from mixed corpora.The survey also recommends checklist-based evaluation to make model capabilities more interpretable and support deeper error analysis.
7 Expanding the Impact of TST
The survey describes how text style transfer can extend beyond standalone generation to support other NLP tasks, specialized applications, and ethical considerations.
- Connecting TST to More NLP Tasks: TST can support paraphrase generation, data augmentation, and adversarial robustness probing.
- Paraphrase Generation: TST and paraphrase generation share content preservation, enabling reuse of models, cross-field method innovation, and related evaluation metrics.
- Data Augmentation and Robustness Probing: TST can generate same-meaning text in different styles for data augmentation and robustness probing.
- Specialized Applications: Specialized applications include persona-consistent dialogue, attractive headlines, style-specific translation, and anonymization.
- Ethical Considerations: The survey frames TST ethics around social impact, privacy, potential fraud, and asking who benefits or is harmed.
- Ethical Considerations: TST may beneficially transform abusive text into normal language, while some gender-transfer uses can enable identity deception or authorship obfuscation.
8 Conclusion
The paper concludes by reviewing deep-learning research on text style transfer and organizing its tasks, evaluation, methods, research agenda, applications, and ethical considerations.
- The survey comprehensively reviews deep-learning text style transfer and categorizes the existing literature.
- It covers task formulation, evaluation metrics, and methods for both parallel and non-parallel data.
- It discusses a research agenda and ways to expand TST's impact across tasks, applications, and ethical considerations.
2008. Automatic natural language style
The cited bibliography entries span neural text generation, style transfer, paraphrasing, disentanglement, evaluation, and related learning methods.
- Several entries address unsupervised or non-parallel text style transfer, including unsupervised machine translation, multiple-attribute rewriting, and dual reinforcement learning.
- The references include work on neural conversation, sentence classification, and content-preserving text generation with attribute controls.
- Other cited works connect style transfer with paraphrase generation, attractive headline generation, prompt tuning, and sequence editing.