Source-linked AI summary
Fair Transfer of Multiple Style Attributes in Text
Karan Dabas, Nishtha Madan, Vijay Arya, Sameep Mehta, Gautam Singh, Tanmoy Chakraborty
TL;DR
Style transfer is challenging because style and content are entangled, and many problems lack parallel data. The paper proposes STMS, which uses a back-translation-derived content representation, one decoder, and feedback from multiple style classifiers. The paper reports superior multi-style-transfer performance for its architecture compared with sequential single-style transfer on Yelp data.
Problem
Style transfer is challenging because style and content are entangled, and many problems lack parallel data.
Method
The paper proposes STMS, which uses a back-translation-derived content representation, one decoder, and feedback from multiple style classifiers.
Results
The paper reports superior multi-style-transfer performance for its architecture compared with sequential single-style transfer on Yelp data.
Takeaways & Limitations
Joint feedback-based control can support fairer simultaneous transfer of multiple soft styles while preserving sentence meaning.
Takeaways & Limitations
Style correlations in the datasets can bias the final content and style strength, and multiple classifiers must work together to guide the decoder.
Abstract
from arXiv · showhide
To preserve anonymity and obfuscate their identity on online platforms users may morph their text and portray themselves as a different gender or demographic. Similarly, a chatbot may need to customize its communication style to improve engagement with its audience. This manner of changing the style of written text has gained significant attention in recent years. Yet these past research works largely cater to the transfer of single style attributes. The disadvantage of focusing on a single style alone is that this often results in target text where other existing style attributes behave unpredictably or are unfairly dominated by the new style. To counteract this behavior, it would be nice to have a style transfer mechanism that can transfer or control multiple styles simultaneously and fairly. Through such an approach, one could obtain obfuscated or written text incorporated with a desired degree of multiple soft styles such as female-quality, politeness, or formalness. In this work, we demonstrate that the transfer of multiple styles cannot be achieved by sequentially performing multiple single-style transfers. This is because each single style-transfer step often reverses or dominates over the style incorporated by a previous transfer step. We then propose a neural network architecture for fairly transferring multiple style attributes in a given text. We test our architecture on the Yelp data set to demonstrate our superior performance as compared to existing one-style transfer steps performed in a sequence.
1 Introduction
The paper frames multi-style transfer as controlling several soft stylistic attributes while preserving sentence meaning, addressing limitations of predominantly single-style methods. It proposes studying sequential single-style transfer and a feedback-based architecture for fair multi-style transfer.
- Existing style-transfer research largely targets individual attributes, including sentiment, gender, or author style.These approaches include transfers such as negative-to-positive sentiment and male-to-female writing style.
- Multi-style transfer aims to convert source stylistic attributes into target attributes while preserving the sentence’s style-independent content.The paper considers combinations such as gender and formality rather than changing only one attribute.
- Sequentially applying single-style transfer is ineffective because later styles can unfairly interact with, nullify, or dominate earlier styles.The paper evaluates sequential transfer as one of two approaches to the multi-style problem.
- The proposed architecture uses a back-translation-based latent representation and multiple style classifiers to guide one decoder toward jointly controlled styles.The architecture is presented as an extension of prior back-translation work for multi-style transfer.
2 Related Work
Prior style-transfer work has mainly handled one style at a time, using either parallel or non-parallel data. The paper situates its approach among methods that alter attributes while balancing content preservation and style control, while also addressing data bias.
- Earlier style-transfer methods generally address one style at a time and differ in whether they require parallel sentence pairs.Parallel data pairs equivalent content across source and target styles, whereas non-parallel data contains separate source- and target-style examples.
- Non-parallel methods include lexicon substitution, attribute-marker removal and retrieval, and autoencoder-based style transfer.Reddy and Knight [2016] use substitutability-based word replacement, while Li et al. [2018] report about 43% overall success but do not control the content-style balance.
- Back-translation-based work removes author-specific stylistic features by translating sentences into another language and back.The paper later adopts this normalization idea for its STMS architecture.
- The paper treats bias in style data as a related problem and uses multi-style conversion as a form of source-data debiasing.It connects this strategy to prior work on gender obfuscation and formality modification.
3 Problem Formulation
The paper formulates multi-style transfer as producing text close to a chosen target style tuple while retaining the input content. It focuses on two attributes, represented through style datasets and an arbiter that scores their strengths.
- With n binary style attributes, the formulation contains 2^n possible style combinations.Each attribute contributes two styles, such as male/female for gender.
- For the paper’s two-attribute case, gender and formality require separate datasets containing male/female and formal/informal sentences.The attributes are denoted s1 and s2, with corresponding datasets D1 and D2.
- The arbiter c(x) maps a sentence to two scores between 0 and 1 representing the strengths of its two stylistic attributes.In this setting, the first score measures maleness and the second measures formality.
- Given an input sentence and one of four target style tuples, the model should generate text whose style scores approach the target while preserving the input content.The target tuples are (1, 1), (1, 0), (0, 1), and (0, 0).
4 Methodology
STMS transfers multiple style attributes by normalizing input content, balancing biased training data, and using one decoder with classifier feedback. Its objective combines reconstruction with separate losses for the two style attributes.
- Multiple-style representation: The method represents style fusion across two datasets, such as gender and formality, while maintaining the input sentence’s content.The fusion operation S1&21 denotes adding styles s11 and s21 to the input.
- Style normalization: STMS normalizes sentences through English-to-French translation followed by back-translation before generating multiple styles.The normalized representation is intended to reduce stylistic bias while preserving sentence content.
- Dataset balancing: A bias rectifier balances underrepresented attribute classes by duplicating minority-class sentences when the data distribution is skewed.The Gender dataset is described as heavily biased toward informal sentences, motivating this preprocessing step.
- Feedback-based generation: STMS uses a single decoder and multiple supervised CNN style classifiers whose feedback induces the desired styles in generated outputs.The classifiers cover the two style attributes and provide feedback during decoder training rather than using separate decoders for each style.
- Joint objective: The decoder encodes each sentence as latent content z and minimizes a weighted sum of decoder, first-classifier, and second-classifier losses.The balancing parameters α and β weight the two classifier losses alongside the decoder loss.
5 Experiments and Results
The experiments compare STMS with single-style and sequential baselines on gender and formality transfer, evaluating content preservation alongside style strength. STMS increases both target attributes, though models trade off style strength against content preservation.
- Baseline Models: Baselines include seq2seq, cross-aligned auto-encoder, and back-translation models, adapted with sequential variants for multi-style transfer.These models were designed originally for one-to-one style transfer.
- Dataset Description: Experiments use Yelp-based gender data and Yahoo Answers formality data, with 2.889M sentences per gender class and 81.844K sentences per formality class.The datasets include train, test, development, and classifier splits.
- Evaluation Techniques: The evaluation measures content preservation with cosine similarity and style strength using separate bidirectional LSTM classifiers for gender and formality.The classifiers reached 81.03% accuracy for male/female classes and 79.7% for formal/informal classes.
- Results: STMS increased both male and formal style attributes, outperforming the other models overall on the female-to-male-and-formal conversion.Its male style score was 0.751, female style score 0.249, and formal style score 0.241; sequential seq2seq achieved the highest formal score at 0.357.
- Results: Content preservation was highest for seq2seq at 0.951, followed by sequential seq2seq at 0.937, BST at 0.908, and STMS at 0.866.Sequential CAE had the lowest preservation score, at 0.626.
6 User Study
The user study examines how writing style affects perceived gender bias and how translated reviews are received. Participants generally associated the translated STMS outputs with greater formality and maleness, while 73% reported satisfaction with them.
- User Perception of Author Writing Bias: The study involved 15 humanities graduates or graduate-level students aged 20–45, including 8 men and 7 women.Participants had prior experience raising awareness about gender bias.
- User Perception of Author Writing Bias: Participants reported that an author’s gender and writing style strongly affected how they interpreted reviews.The first study phase asked whether readers inferred the author’s gender and then perceived the review accordingly.
- User Evaluation of Our System: The follow-up study asked participants to compare outputs for satisfaction, content preservation, formality, and perceived maleness and formalness.The same subjects evaluated the translated versions after the initial bias-perception survey.
- User Evaluation of Our System: 11 of 15 participants judged STMS outputs to reflect more maleness and some formalness, compared with 4 favoring BST.The source sentences were female in style, so the comparison assessed whether the translations conveyed the desired target attributes.
- User Evaluation of Our System: 73% of participants were satisfied with STMS translations, and 65% felt better after reading the translated reviews.All participants selected seq2seq as having the highest content preservation, while a majority judged STMS more formal.
7 Conclusion and Future Work
The paper formulates multi-style transfer as preserving sentence meaning while jointly controlling multiple stylistic features, and proposes sequential and jointly guided approaches. Experiments indicate both models strengthen target styles while retaining content, while dataset bias and classifier coordination remain important challenges.
- 7 Conclusion and Future Work: Both Sequential seq2seq and STMS improve target style strength while retaining the underlying content across multi-style transfer experiments.The evaluation used gender- and formality-related datasets and included a user study.
- 7 Conclusion and Future Work: The proposed approaches combine individual style models in series or use back-translation with feedback from multiple style classifiers to guide decoding jointly.
- 7 Conclusion and Future Work: Mutual bias between style datasets can affect final content and style strength, requiring classifier-loss weights to account for those biases.The paper observes that gender data are often informal, which can leave outputs highly informal when converting to male and formal styles.
- 7 Conclusion and Future Work: Non-parallel style datasets make it challenging to build classifiers trained on one dataset that work accurately on other datasets.