PaperlayerView exact revision on arXiv

Source-linked AI summary

PowerTransformer: Unsupervised Controllable Revision for Biased Language Correction

Xinyao Ma, Maarten Sap, Hannah Rashkin, Yejin Choi

arXiv:2010.13816v1cs.CLcs.AI

TL;DR

Narratives and media can reflect biased character portrayals, and parallel data for correcting them is unavailable. The paper introduces POWERTRANSFORMER, which uses agency connotation frames and reconstruction and paraphrase objectives; it reports stronger results than related methods and a movie-script case study that increases female characters’ agency.

  • Problem

    Narratives and news can reflect societal stereotypes, while controllable debiasing must work without parallel data and make purposeful, precise revisions.

    Source: §1 · pp. 1–2

  • Method

    POWERTRANSFORMER uses reconstruction and paraphrase objectives and steers revisions with agency connotation frames, training-time control tokens, and generation-time vocabulary boosting.

    Source: §1 · p. 2

  • Results

    POWERTRANSFORMER outperforms stylistic-rewriting methods on human and automatic metrics; ablations report 47% gain in accuracy for the joint objective and 12% gain in accuracy for agency scaling.

    Source: §1 · p. 2

  • Takeaways & Limitations

    In movie scripts, revisions significantly increase female characters’ agency levels, demonstrating a case-study application of controllable debiasing.

    Source: §1 · p. 2, §7 · p. 9

    The movie-data analysis cannot support more inclusive gender analyses, such as non-binary categories, because information about individuals is limited.

    Source: §5.2 · p. 8

Abstract

from arXiv · show

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.

1 Introduction

The introduction frames controllable debiasing as targeted revision of biased character portrayals, using agency connotation frames to guide edits. POWERTRANSFORMER combines reconstruction and paraphrase objectives with agency controls, and is evaluated on sentence rewriting and movie scripts.

1 Introduction

  • Narratives and news can reflect societal stereotypes, and targeted rewriting could help reshape character portrayals in media.The introduction illustrates this with changing a character’s portrayal from daydreaming about becoming a doctor to pursuing that dream.Source: §1 · p. 1
  • Effective debiasing must revise overt action choices and subtler framing, while avoiding unnecessary meaning changes and learning without parallel data.The introduction argues that surface-level paraphrasing may leave underlying biased events uncorrected.Source: §1 · pp. 1–2
  • Controllable Debiasing targets implicit, potentially unwanted character bias through revisions guided by connotation frames of power and agency.These frames encode implied power and agency levels associated with predicates.Source: §1 · p. 2
  • POWERTRANSFORMER uses reconstruction and paraphrase objectives, with control tokens during training and agency-based vocabulary boosting during generation.The approach addresses the lack of parallel supervised data.Source: §1 · p. 2
  • POWERTRANSFORMER outperforms stylistic-rewriting methods on human and automatic metrics; ablations report 47% gain in accuracy for the joint objective and 12% gain in accuracy for agency scaling.The model also makes minimal changes to sentence meaning.Source: §1 · p. 2
  • In movie scripts, revisions significantly increase female characters’ agency levels, thereby reducing gender bias.The paper presents this as a case study using a corpus of modern English movies.Source: §1 · p. 2

2 Controllable Debiasing

Controllable Debiasing formalizes stylistic revision to debias character portrayals while balancing meaningful changes with precision. It focuses on the power and agency conveyed by actions, with potential use in writing systems.

2 Controllable Debiasing

  • Controllable Debiasing may change the underlying meaning of events to avoid merely paraphrasing bias, but revisions should avoid unnecessary meaning changes.This contrasts with tasks such as politeness transfer, where full meaning preservation is required.Source: §2 · p. 2
  • Connotation frames capture implicit relations between a verb, its agent, and its theme; agency means the capacity to intentionally change or act upon one’s environment.The paper uses positive, neutral, and negative agency dimensions.Source: §2 · p. 2
  • Writing systems using models trained for this task could help authors portray characters in less biased ways and mitigate stereotypical portrayals’ negative effects.The passage names news, stories, and movies as potential settings.Source: §2 · p. 2

3 POWERTRANSFORMER

POWERTRANSFORMER trains jointly on reconstruction and paraphrasing, then steers agency during decoding. Its stated scope centers agency, while other bias operationalizations are left for future work.

3 POWERTRANSFORMER

  • POWERTRANSFORMER jointly learns to reconstruct partially masked story sentences and paraphrase external examples, addressing scarce parallel supervision while adding generation-time vocabulary boosting for fine-grained agency control.Source: §3 · pp. 2–3, Figure 2 · p. 3
  • The method operationalizes portrayal bias through agency; exploring power or alternative bias framings is left for future work.Source: §3 · p. 2

3.1 Model Overview

  • The encoder receives BPE input scrubbed of agency markers plus one of three target-agency tokens: <Pos>, <Equal>, or <Neg>.Source: §3.1 · p. 3

3.2 Joint Objective

  • Training combines reconstruction of sentences with agency-indicative verbs masked and paraphrasing from an external corpus, conditioning both tasks on masked input and target agency.Source: §3.2 · p. 3

3.3 Controlled Decoding with Vocab Boosting

  • At decoding, vocabulary boosting rescales token logits using agency embeddings, a target-agency one-hot vector, and strength β to increase the likelihood of matching-agency words.Source: §3.3 · p. 4

4 Controllable Debiasing Experiments

Experiments assess agency control, meaning preservation, and output quality through automatic metrics and human judgments. POWERTRANSFORMER performs better overall than related rewriting baselines, with trade-offs across agency, meaning preservation, and fluency.

4 Controllable Debiasing Experiments; 4.3 Experimental Setup

  • The experiments combine automatic comparisons of the full model and ablations, related-work baselines, and crowdsourced judgments.Decoding uses top-p=0.4 nucleus sampling and boosting strength β=5.Source: §4 · p. 4, §4.3 · p. 5

4.1 Datasets; 4.2 Metrics

  • Training uses agency-balanced ROC stories and subtitle paraphrases; evaluation measures agency change, semantic similarity, fluency, repetition, and output diversity.Source: §4.1 · p. 4, §4.2 · p. 4

4.4.1 Ablated Baselines; 4.5.1 Baselines

  • Ablations compare joint reconstruction and paraphrase training with paraphrase-only training, agency vocabulary boosting, and a retrieved-verb variant.The external baselines are BST and PPLM, evaluated only on positive and negative agency subsets because they were designed for binary style transfer.Source: §4.4.1 · p. 5, §4.5 · p. 5

4.4.2 Results

  • Joint+Boost produces the most accurate target agency and greatest meaning preservation, while increasing output diversity with marginally more repeated bigrams.The ablation results are reported on the development set across agency, meaning, fluency, repetitiveness, and diversity metrics.Source: §4.4.2 · p. 5, Table 2 · p. 5
  • SupplyVerb does not improve agency or meaning metrics and worsens fluency; retrieved verbs can reflect a different sense and break grammaticality.Source: §4.4.2 · p. 5

4.5.2 Results

  • POWERTRANSFORMER performs better overall than BST and PPLM, though BST has slightly higher agency accuracy with lower diversity and meaning preservation, while PPLM shows the opposite pattern.The comparison covers negative-to-positive and positive-to-negative agency revisions on the test set.Source: §4.5.2 · p. 6, Table 3 · p. 6

4.6.1 Human Evaluation Task

  • Human evaluation compares three revisions per input-target pair, with three judges rating gibberish, target agency, and meaning preservation.The evaluation uses a random subset of 100 test examples and compares POWERTRANSFORMER with PPLM, BST, and Joint+noBoost.Source: §4.6.1 · pp. 6–7

4.6.2 Results

  • Human raters preferred POWERTRANSFORMER revisions over all baselines for target agency; for meaning preservation, it was always preferred over BST, less decisively over PPLM, and preference reversed against the ablated noBoost model.BST revisions were marked as gibberish substantially more than those by other models (63% vs. 3-7%).Source: §4.6.2 · p. 7

5 Gender Bias in Movies

The study applies PowerTransformer to rebalance female characters’ agency in movie-script portrayals and examines whether revision reverses gendered agency patterns. Its evidence is limited by binary gender inference and a focus on movie scripts and agency-based portrayals.

5.1 Movie Scripts Corpus

  • Across 16,763 characters in 767 movies, 68% were inferred men and 32% women; female characters also had fewer narrations, words, and verbs.On average, women appeared in 27 versus 34 narrations, 329 versus 435 words, and 41 versus 54 verbs; each reported difference was significant.Source: §5.1 · p. 8

5 Gender Bias in Movies

  • As a proof of concept, the study revises only sentences describing female characters, increasing positive-agency verbs and decreasing negative-agency or passive verbs.Figure 4 compares average agency-verb levels for female characters in original and revised scripts.Source: §5.2 · p. 8, Figure 4 · p. 8

5.2 Debiasing Portrayal in Movies

  • Gender assignment is binary, and 2,597 characters had no inferable gender; limited individual information prevents non-binary analyses.Source: §5.2 · p. 8
  • Revisions reverse the movie-script gender association: women show significantly more positive agency (β′pos = −62.6, p < 0.001) and less negative agency (β′neg = 8.7, p < 0.001).In original scripts, men had significantly higher positive agency (βpos = 1.2, p < 0.001) and lower negative agency (βneg = −0.3, p < 0.001) than women.Source: §5.2 · p. 8
  • The findings concern movie-script portrayals; the authors identify other domains and alternative views of portrayal bias as directions for future work.Suggested extensions include news or fiction, regard, and broader views of gender roles such as masculine-default cultures.Source: §5.2 · p. 8

6 Related Work

Related work situates Controllable Debiasing within unsupervised stylistic rewriting while distinguishing its focus on social bias and its requirement to revise an input sentence. PowerTransformer also adds an out-of-domain paraphrasing objective to a modular rewriting approach.

6 Related Work

  • Controllable Debiasing formalizes unsupervised stylistic rewriting for social bias, contrasting with supervised approaches that benefit from parallel corpora.Unsupervised alternatives include encoder-decoder systems with discriminators and backtranslation; PowerTransformer adds an out-of-domain paraphrasing objective to Li et al. (2018a).Source: §6 · pp. 8–9
  • Unlike offensiveness-oriented rewriting (Nogueira dos Santos et al., 2018), Controllable Debiasing targets social biases through nuanced power-and-agency connotations rather than binary offensiveness.Source: §6 · p. 9
  • Controllable revision differs from controllable generation by requiring revision of an input sentence as well as generation with desired attributes.Source: §6 · p. 9

7 Conclusion

The paper introduces Controllable Debiasing to revise character portrayals through power and agency connotation frames. PowerTransformer outperforms ablations and baselines in automatic and human evaluations, and a movie-script case study demonstrates feasibility.

7 Conclusion

  • Controllable Debiasing targets implicit, potentially unwanted bias in character portrayals using power and agency connotation frames.Source: §7 · p. 9
  • PowerTransformer outperforms ablations and baselines on automatic and human evaluations.Source: §7 · p. 9
  • A movie-script case study demonstrates the feasibility of using Controllable Debiasing to debias character portrayals.Source: §7 · p. 9

A Additional data description

The ROC story corpus comprises everyday-scenario stories, filtered by agency determinations and sentence verb counts for this task.

A.1 ROC story corpus

  • The source corpus contains 100,000 five-sentence crowdworker stories about realistic everyday scenarios.Source: §A.1 · p. 12
  • Sentences with indeterminable agency levels are excluded when selecting data for the task.Source: §A.1 · p. 12
  • Sentences with four or more verbs are filtered out to avoid masking too many content words.Source: §A.1 · p. 12

A.2 Paraphrase corpus

The paraphrase corpus uses spoken-dialogue paraphrases from movie and TV subtitles, with OpusParcus pairs selected and filtered for the task.

A.2 Paraphrase corpus

  • The corpus contains paraphrases of spoken dialogue extracted from movie and TV subtitles.Source: §A.2 · p. 12
  • OpusParcus subtitle sentences were automatically aligned using probabilistic metrics, including roundtrip-translation-model likelihood and pointwise mutual information.Source: §A.2 · p. 12
  • The English OpusParcus training corpus received the ROC filtering, and the top 10% of paraphrases by original-paper PMI scores were selected.Source: §A.2 · p. 12
  • Selected pairs were chosen for roughly equal representation across agency-level pairs, then leading periods and commas were stripped.Source: §A.2 · p. 12

B Experimental details

The experimental setup uses a Hugging Face implementation of GPT and reports optimizer, batch, sequence-length, and steering settings.

B Experimental details

B.1 POWERTRANSFORMER details

The appendix records training and tuning choices for PowerTransformer and related baselines, describes a PPLM adaptation, and documents movie-script character extraction and evaluation materials.

B.1 POWERTRANSFORMER details

  • POWERTParaOnly+None and POWERTJoint+None train for 10 epochs at a learning rate of 1e-5 with AdamW; each epoch takes about an hour.For POWERTParaOnly+None, the learning rate was manually tuned over [1e-6, 1e-3] seven times, and nucleus-sampling p was set to 0.4 after trying five values in [0.4, 0.9].Source: §B.1.1 · p. 12, §B.1.3 · p. 12
  • The Static variants load their corresponding None models and add logit rescaling, tuning β over [0, 10] and using 5 in the final model.The same p as POWERTParaOnly+None is used for both Static variants.Source: §B.1.2 · p. 12, §B.1.4 · p. 12

B.2 PPLM details

  • PPLM replaces its pretrained-language-model base with a GPT-2 denoising autoencoder trained on ROC-only data to reconstruct input sentences.At decoding, inputs update toward meaning-preserving sentences favored by an agency discriminator; overly strong target-label settings can produce degenerate text.Source: §B.2 · pp. 12–13

B.3 Backtranslation details

  • The backtranslation baseline translates lowercase positive- and negative-agency examples to French, trains an agency classifier and two decoders, and selects decoder settings by dev-set word-level accuracy.It uses SGD with batch size 64; the classifier uses a learning rate of 5 for 6 epochs, while decoder settings are selected from a grid search.Source: §B.3 · p. 13

C.1 Extracting gender from characters

  • Movie-script characters are identified from all-caps mentions and cross-referenced against gendered names and gendered words.The text also describes assigning sentences to characters when their names appear, and splitting narratives for rewriting.Source: §C.1 · p. 13, §B.2 · p. 13
  • The appendix shows human-evaluation annotation and qualification tasks, notes that workers rated three sentence pairs in the real qualification task, and provides dev-set revision examples.The qualification screenshot displays only one of the three pairs.Source: Figure 5 · p. 14, Figure 6 · p. 15, Table 6 · p. 16

Related papers

Loading 2010.13816v1…