PowerTransformer: Unsupervised Controllable Revision for Biased Language Correction

Xinyao Ma, Maarten Sap, Hannah Rashkin, Yejin Choi

Introduction

Language often harbors implicit social biases through subtle framing, even in the absence of overt slurs. In narrative contexts, characters are frequently portrayed with varying levels of "agency"—the capacity to act intentionally. Research indicates that female characters are often described with low-agency verbs (e.g., "waited," "daydreamed"), whereas male characters are associated with high-agency actions (e.g., "pursued," "decided").

Addressing this "portrayal bias" is challenging due to the lack of parallel "biased-to-unbiased" datasets and the need to preserve narrative meaning while altering pragmatic framing. PowerTransformer is introduced as an unsupervised framework to rewrite text, transforming character agency without requiring direct supervision.

PowerTransformer revises sentences to alter character agency while preserving core meaning.

Methodology

PowerTransformer utilizes an encoder-decoder architecture based on OpenAI's GPT. The framework operates without parallel data by integrating three key components:

1. Connotation Frames

The model leverages a predefined lexicon that assigns agency scores to verbs. This allows the system to identify the pragmatic power dynamics inherent in specific actions.

2. Joint Training Objective

The model is trained simultaneously on two tasks:

  • Reconstruction (Denoising): The model learns to fill in masked verbs given a "target agency" token (e.g., <Pos> for high agency), effectively learning the relationship between context and agency.
  • Paraphrasing: By training on an external paraphrase corpus, the model learns to maintain semantic consistency during the revision process.

3. Vocab Boosting at Inference

A novel "Vocab Boosting" technique is applied during decoding. Instead of relying solely on learned weights, the model increases the logit scores of verbs that align with the desired agency level. This steers the generation toward appropriate "high-power" or "low-power" words that fit the grammatical and narrative context.

The PowerTransformer architecture combines joint training with vocabulary boosting at decoding time.

Datasets and Evaluation

The model was trained and tested using several diverse datasets:

  • ROC Story Corpus: 100,000 five-sentence stories for the reconstruction task.
  • OpusParcus: A large paraphrase dataset derived from movie and TV subtitles.
  • Movie Scripts Corpus: 767 modern English movie scripts used to evaluate the model's ability to mitigate real-world gender bias.

Key Results

PowerTransformer demonstrates significant improvements over baseline models:

  • Target Accuracy: The model achieved 89% accuracy in reaching the target agency level, far surpassing PPLM (13%).
  • Meaning Preservation: It maintained a high 0.96 BERTScore, ensuring that revisions remained faithful to the original story logic.
  • Human Preference: In head-to-head evaluations, human judges preferred PowerTransformer’s revisions for both agency-matching and semantic preservation.
  • Bias Mitigation: When applied to movie scripts, the model successfully adjusted female character portrayals to agency levels comparable to male characters.

Limitations and Future Work

While effective, the model's reliance on a fixed lexicon limits its performance on out-of-vocabulary verbs. Additionally, the current implementation uses coarse agency categories (Positive, Neutral, Negative) and focuses on binary gender portrayals. Future research may explore more granular social dimensions and broader social constructs.

Summary updated

Abstract

Unconscious biases continue to be prevalent in modern text and media, calling for algorithms that can assist writers with bias correction. For example, a female character in a story is often portrayed as passive and powerless (“_She daydreams about being a doctor_”) while a man is portrayed as more proactive and powerful (“_He pursues his dream of being a doctor_”). We formulate **Controllable Debiasing**, a new revision task that aims to rewrite a given text to correct the implicit and potentially undesirable bias in character portrayals. We then introduce PowerTransformer as an approach that debiases text through the lens of connotation frames (Sap et al., 2017), which encode pragmatic knowledge of implied power dynamics with respect to verb predicates. One key challenge of our task is the lack of parallel corpora. To address this challenge, we adopt an unsupervised approach using auxiliary supervision with related tasks such as paraphrasing and self-supervision based on a reconstruction loss, building on pretrained language models. Through comprehensive experiments based on automatic and human evaluations, we demonstrate that our approach outperforms ablations and existing methods from related tasks. Furthermore, we demonstrate the use of PowerTransformer as a step toward mitigating the well-documented gender bias in character portrayal in movie scripts.
BibTeX
@inproceedings{ma2020powertransformer,
  author = {Xinyao Ma and Maarten Sap and Hannah Rashkin and Yejin Choi},
  title = {{PowerTransformer: Unsupervised Controllable Revision for Biased Language Correction}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  doi = {10.18653/v1/2020.emnlp-main.602},
}
Figure