PaperlayerView exact revision on arXiv

Source-linked AI summary

Reformulating Unsupervised Style Transfer as Paraphrase Generation

Kalpesh Krishna, John Wieting, Mohit Iyyer

arXiv:2010.05700v1cs.CL

TL;DR

Style transfer aims to change style without changing meaning, but existing systems can fail to preserve semantic content. The paper reformulates unsupervised style transfer as paraphrase generation using automatically created data, and reports stronger evaluations alongside a critique of metrics and a large multi-style benchmark.

  • Problem

    Style transfer should preserve semantics, but attribute-transfer systems can alter semantic properties, and fewer than 25% of outputs from two state-of-the-art formality systems were rated as paraphrases.

  • Method

    STRAP creates pseudo-parallel paraphrase data, trains style-specific inverse paraphrasers, and uses them to generate a requested style without parallel cross-style data.

  • Results

    STRAP significantly outperforms state-of-the-art systems on formality transfer and Shakespeare imitation, by 2-3x on automatic evaluations and 4-5x on human evaluations.

  • Takeaways & Limitations

    The paper argues that paraphrase-based modeling better captures semantic preservation, while its survey exposes weaknesses in existing automatic style-transfer evaluation.

  • Takeaways & Limitations

    Baseline models were not readily adapted to 11-way style transfer without significant code or model modifications, given their formulations beyond two styles.

Abstract

from arXiv · show

Modern NLP defines the task of style transfer as modifying the style of a given sentence without appreciably changing its semantics, which implies that the outputs of style transfer systems should be paraphrases of their inputs. However, many existing systems purportedly designed for style transfer inherently warp the input's meaning through attribute transfer, which changes semantic properties such as sentiment. In this paper, we reformulate unsupervised style transfer as a paraphrase generation problem, and present a simple methodology based on fine-tuning pretrained language models on automatically generated paraphrase data. Despite its simplicity, our method significantly outperforms state-of-the-art style transfer systems on both human and automatic evaluations. We also survey 23 style transfer papers and discover that existing automatic metrics can be easily gamed and propose fixed variants. Finally, we pivot to a more real-world style transfer setting by collecting a large dataset of 15M sentences in 11 diverse styles, which we use for an in-depth analysis of our system.

1 Introduction

The paper argues that style transfer should change lexical and syntactic style while preserving sentence semantics, unlike attribute transfer that can alter semantic content. It introduces STRAP, an unsupervised paraphrase-based method, and reports strong evaluation and dataset contributions.

  • Style transfer should modify a sentence’s style while preserving its semantics, so outputs should be paraphrases of their inputs.
  • Attribute transfer can modify attribute-specific content words, such as sentiment-bearing terms, thereby warping both stylistic and semantic properties.
  • STRAP reformulates unsupervised style transfer as controlled paraphrase generation without requiring parallel data between styles.
  • STRAP creates pseudo-parallel data with a diverse paraphraser, trains style-specific inverse paraphrasers, and swaps inverse models at test time for transfer.
  • STRAP outperforms state-of-the-art systems by 2-3x on automatic evaluations and 4-5x on human evaluations for formality transfer and Shakespeare imitation.
  • The paper surveys 23 style transfer papers and introduces CDS, a dataset of 15M English sentences spanning 11 diverse styles.

2 Style Transfer via Paraphrasing

The paper motivates style transfer through semantic preservation and reformulates it as controlled paraphrase generation. STRAP normalizes sentences with diverse paraphrasing, then reconstructs them in a selected style using style-specific inverse paraphrasers.

  • Fewer than 25% of outputs from two state-of-the-art formality-transfer systems were judged to be paraphrases of their inputs.
  • STRAP uses raw text from distinct styles and creates pseudo-parallel pairs because parallel sentences across styles are unavailable.
  • A diverse paraphraser normalizes input sentences by removing information predictive of their original style.
  • Style-specific inverse paraphrase models reconstruct original sentences from normalized paraphrases, learning to insert salient features of the original style.
  • At inference, STRAP normalizes an arbitrary sentence and then applies the inverse paraphraser for the desired target style.
  • The system fine-tunes pretrained GPT-2 models for the paraphraser and separate inverse paraphrasers, improving fluency and generalization to small style-specific datasets.
  • Output diversity is promoted by aggressively filtering PARANMT-50M, because maximizing lexical and syntactic diversity improves style normalization.

3 Evaluating style transfer

The paper argues that existing style-transfer evaluations inadequately measure joint transfer accuracy, semantic similarity, and fluency, and introduces sentence-level aggregation to address this problem. It also replaces weak metric implementations and shows that corpus-level aggregation can be gamed.

  • Evaluation methodology: 23 papers reveal widespread evaluation weaknesses, including poorly defined quality measures and misleading aggregation of accuracy, similarity, and fluency.The paper proposes improved implementations for these evaluation dimensions.
  • Evaluation methodology: 14 of 23 surveyed papers use 1-layer CNN classifiers for transfer accuracy, despite RoBERTa-large achieving higher test accuracy on both datasets.RoBERTa-large reaches 90.4% versus 83.5% on Shakespeare and 94.8% versus 92.4% on Formality.
  • Evaluation methodology: 15 of 23 papers use n-gram metrics such as BLEU for semantic similarity, despite unreliable correlations between n-gram overlap and human similarity judgments.The paper motivates evaluating semantic preservation more directly.
  • Evaluation methodology: Only 10 of 23 papers evaluate fluency, with 9 using perplexity, which is unbounded and can favor unnatural sentences containing common words.The paper replaces perplexity with a RoBERTa-large classifier trained on CoLA acceptability judgments.
  • Human evaluation: Human evaluation measures similarity and fluency jointly with a three-point paraphrase scale, but excludes human accuracy judgments because crowdsourced workers may lack familiarity with target styles.The authors therefore use classifier-based accuracy evaluation.
  • Aggregation of metrics: Corpus-level aggregation can reward systems that optimize only a subset of metrics, such as copying for similarity or repeating a target-style sentence for accuracy and fluency.These metrics are inversely correlated, making aggregation especially difficult.
  • Aggregation of metrics: A naïve copy-or-random-retrieval baseline beats state-of-the-art systems under geometric-mean corpus aggregation despite performing no style transfer.The baseline copies with probability p and otherwise retrieves a random target-style sentence.
  • Aggregation of metrics: The proposed J(·) metric combines sentence-level ACC, SIM, and FL before averaging, assigning zero to incorrectly classified or disfluent outputs.This prevents separate metric strengths from compensating for failures on individual sentences.

4 Experiments & Results

The experiments evaluate STRAP on Shakespeare author imitation and formality transfer while training only on non-parallel data. STRAP substantially outperforms prior systems on automatic and human evaluations, especially in semantic similarity and fluency.

  • Datasets and setup: STRAP is evaluated on Shakespeare author imitation and formality transfer, using parallel data only for automatic evaluation and non-parallel splits for training.The Shakespeare dataset contains 37k training sentences, while Formality contains 105k sentences.
  • Baselines: The comparison includes COPY, NAÏVE, REF, UNMT, and DLSM, with DLSM identified as the prior state of the art on both datasets.NAÏVE randomly copies the input or retrieves a target-style sentence; REF evaluates reference sentences.
  • Automatic evaluation: 45.5 vs 18.6 on Formality and 34.7 vs 16.3 on Shakespeare are STRAP’s automatic J(·) scores against DLSM.The comparisons report substantial gains over the prior state of the art.
  • Human evaluation: 41.3 vs 10.0 on Formality and 47.3 vs 9.3 on Shakespeare are STRAP’s human-evaluation scores against DLSM for SIM and FL.The paper describes these as larger improvements than those observed in automatic evaluation.
  • Results: STRAP improves SIM and FL across evaluations while maintaining similar ACC relative to prior systems.The authors report this pattern across the evaluated datasets.

5 Ablation studies

Ablations show that STRAP’s gains depend on paraphrase diversity, pretrained language models, and paraphrasing during inference, while separate inverse paraphrasers outperform a single style-conditional model.

  • Paraphrase diversity: Removing diversity filtering lowers ACC and raises SIM because the paraphraser copies more of the input, reducing the overall J(·) score.This supports diversity as an important component of the paraphrase generation step.
  • Inference-time paraphrasing: 45.5 to 20.7 in Formality and 34.7 to 23.3 in Shakespeare: removing paraphrasing during inference sharply reduces ACC and the overall score.SIM and FL remain largely unaffected, supporting the hypothesis that paraphrasing normalizes the input.
  • Language-model pretraining: Replacing GPT-2 with LSTM seq2seq models causes large drops in SIM and FL, making performance significantly worse than STRAP.The LSTM model remains competitive with UNMT and DLSM on the aggregate Formality score, but not with STRAP.
  • Decoding strategy: Higher nucleus-sampling p improves diversity but trades off increased ACC for lowered SIM, with higher values degrading J(·) performance.The experiments find p = 0.6 similar to greedy decoding on J(·) metrics.
  • Model parameterization: A single style-conditional inverse paraphraser is less effective than training separate inverse paraphrasers per style, primarily because of lower ACC.The authors hypothesize that one model faces conflicts between different style distributions.

6 Towards Real-World Style Transfer

CDS extends style-transfer evaluation to 11 diverse styles and 110 potential transfer directions, enabling analysis beyond two-style benchmarks. STRAP normalizes stylistic features through diverse paraphrasing and achieves measurable transfer performance, while retaining known generation failures.

  • Corpus of Diverse Styles: CDS contains 11 diverse styles and 15M tokens, creating 110 potential transfer directions instead of the two-style setup common in prior benchmarks.The dataset includes sources such as Shakespeare, tweets, romantic poetry, and conversational speech.
  • Diverse paraphrasing normalizes stylized text: 88.9% classifier accuracy on original CDS validation data fell to 42.5% after diverse paraphrasing, indicating substantial reduction in recognizable stylistic features.The non-diverse paraphraser produced a smaller drop, to 51.5%.
  • Style Transfer on CDS: 48.4% transfer accuracy, a paraphrase similarity score of 63.5, and 71.1% fluent generations yielded a J(ACC,SIM,FL) score of 20.7 on CDS.STRAP transferred 1K sampled sentences from each style to each of the other 10 styles.
  • Controlled comparisons: In controlled comparisons on Shakespeare and English Tweets, STRAP greatly outperformed UNMT and DLSM, especially on SIM and FL.Baseline models often produced arbitrary target-style features while ignoring input semantics, leading to poor SIM despite high ACC.
  • Qualitative examples: STRAP’s common failures include poor handling of rare words, dropped or modified semantic content, and hallucinations reflecting the training distribution.These errors resemble failures in contemporary text-generation systems.

7 Related Work

Prior unsupervised style-transfer work uses disentanglement, policy gradients, retrieval, backtranslation, and Transformer-based methods. The paper instead emphasizes automatically learned style-specific paraphrasers without inference-time over-generation.

  • Existing approaches: Unsupervised style transfer has been modeled with attribute-based disentanglement, policy-gradient training, and retrieval-based approaches.Backtranslation has also been used to model semantic preservation, though prior work shows it can warp semantics.
  • Comparison scope: The baseline models were not readily extended to 11-way transfer because their formulations were complex beyond two styles or required separate models for 110 directions.This limitation affected controlled comparison on the full CDS setting.
  • Backtranslation: The paper uses backtranslation only to build paraphraser training data after heavy filtering, rather than directly as the complete style-transfer method.This design addresses the reported risk of semantic warping in backtranslation.
  • Related paraphrase methods: Unlike a related handcrafted pipeline, the paper automatically learns style-specific paraphrasers and avoids over-generation during inference.The comparison is made with Grӧndahl and Asokan’s paraphrase over-generation and filtering approach.

8 Conclusion

The paper reframes style transfer as controlled paraphrase generation and introduces STRAP, evaluation critiques, and CDS. It concludes with diverse paraphrasing and broader style coverage as central directions for future work.

  • Contributions: The paper models style transfer as controlled paraphrase generation and presents a simple unsupervised method using diverse paraphrasing.Its contributions also include evaluation fixes and a large multi-style dataset.
  • Evaluation: A survey of 23 papers critiques current style-transfer evaluation and proposes fixes to common shortcomings.The conclusion identifies evaluation methodology as a separate contribution from the transfer model.
  • Dataset: The paper collects a dataset containing 15M sentences from 11 diverse styles.CDS broadens the empirical setting beyond conventional two-style benchmarks.
  • Future work: Future work includes exploring diverse paraphrasing for applications such as data augmentation and performing style transfer across broader settings.The supplied conclusion explicitly names data augmentation as one future application.

A.1 PARANMT-50M Filtering Details

The paraphrase model is trained from filtered PARANMT-50M backtranslated pairs, with filters targeting semantic preservation, lexical diversity, and English-language quality. Filtering is applied as a pipeline whose corpus sizes are tracked after each stage.

  • Data source: The paraphrase model is trained in a sequence-to-sequence fashion on PARANMT-50M, a large corpus created by backtranslating Czech text.The corpus is described as large and noisy before aggressive filtering.
  • Content filtering: Pairs scoring below 0.5 on a paraphrase similarity model are removed, and paired sentences may differ by at most 5 words.Remaining pairs must have average token length between 7 and 25.
  • Lexical diversity filtering: Pairs with more than 50% unigram or trigram overlap are filtered out to encourage lexical diversity and discourage copying.This filtering targets diversity in automatically generated paraphrases.
  • Language identification filtering: Sentence pairs classified as non-English on both sides are discarded using langdetect.The filtering stage is an English-language quality check.
  • Filtering pipeline: Filtering is pipelined, and PARANMT-50M corpus size is reported after each filtering stage.The staged effects are summarized in Table 8.
  • Diversity measurement: Word-order diversity is measured by shuffling percentage, where identical ordering is 0% shuffled and reverse ordering is 100% shuffled.This convention supports interpretation of the lexical and syntactic diversity results.

A.2 Generative Model Details

The paper uses decoder-only GPT-2 in an encoder-free sequence-to-sequence setup for both paraphrasing and style-specific inverse paraphrasing. Inputs and outputs are concatenated with special tokens and segment information, then outputs are generated autoregressively.

  • GPT-2 is fine-tuned for both the paraphrase model and the style-specific inverse paraphrase model.
  • The encoder-free setup feeds input and output sequences to one decoder, separated by a special token and distinguished with segment embeddings.
  • Architecture: Training predicts output tokens with next-word prediction and cross-entropy loss after concatenating the input sequence with the output prefix.
  • Architecture: During inference, GPT-2 receives the input and beginning-of-sentence token, then generates tokens autoregressively until the end-of-sentence token.
  • Architecture: The model uses shared token embeddings plus learnable positional and segment embeddings, with segments indicating whether tokens belong to the input or output.
  • Other seq2seq alternatives: The algorithm is agnostic to the sequence-to-sequence architecture, while the authors identify T5 and BART as future alternatives to the encoder-free approach.

A.3 Classifier Model Details

The appendices describe classifier and comparison experiments, including GPT-2 and LSTM baselines, evaluation-pipeline comparisons, and analysis of paraphrase diversity and CDS style transfer. They also identify metric and fluency-classifier limitations.

  • More Comparisons with Prior Work: The authors compare prior systems with their formality-transfer model using outputs from a public repository, but note a training–evaluation domain shift between corpus splits.
  • More Comparisons with Prior Work: 53.1% ACC, 55.2 SIM, and 56.5% FL yield a J(A,S,F) score of 18.4 for Sudhakar et al. on Shakespeare.
  • More Comparisons with Prior Work: 41.7% ACC, 67.8 SIM, and 67.7% FL yield a J(A,S,F) score of 18.1 for Sudhakar et al. on Formality.
  • More Comparisons with Prior Work: The Sudhakar et al. models produced little abstraction, near-copying, and low fluency, with deletion and generation identified as a possible source of the fluency problem.
  • Diverse Paraphrasing on CDS: The diverse paraphraser produced 51% unigram change and 27% word shuffling, versus 28% and 6% for the non-diverse paraphraser, with SIM=72.5 versus 83.9.
  • CDS Evaluation: For CDS, Switchboard, Lyrics, and James Joyce were among the styles transferred more successfully, while nucleus sampling traded ACC against SIM and p = 0.6 was optimal for ACC.
  • CDS Evaluation: Fluency scores on CDS may be misleading because some original styles are classified as disfluent by a CoLA-trained classifier.
  • A Survey of Evaluation Methods: Only 3 out of 23 surveyed papers used absolute sentence-level aggregation, while other approaches used relative, flawed, or absent aggregation procedures.
Loading 2010.05700v1…