Gender and Representation Bias in GPT-3 Generated Stories

Using topic modeling and lexicon-based word similarity, we find that stories generated by GPT-3 exhibit many known gender stereotypes. Generated stories depict different topics and descriptions depending on GPT-3’s perceived gender of the character in a prompt, with feminine characters more likely to be associated with family and appearance, and described as less powerful than masculine characters, even when associated with high power verbs in a prompt. Our study raises questions on how one can avoid unintended social biases when using large language models for storytelling.

[1]  Rachel Rudinger,et al.  “You Are Grounded!”: Latent Name Artifacts in Pre-trained Language Models , 2020, EMNLP.

[2]  T. Underwood,et al.  The Quiet Transformations of Literary Studies: What Thirteen Thousand Scholars Could Tell Us , 2014 .

[3]  Melody Dye,et al.  Gender bias at scale: Evidence from the usage of personal names , 2019, Behavior research methods.

[4]  Noah A. Smith,et al.  Creative Writing with a Machine in the Loop: Case Studies on Slogans and Stories , 2018, IUI.

[5]  Eva Flicker.  Between Brains and Breasts—Women Scientists in Fiction Film: On the Marginalization and Sexualization of Scientific Competence , 2003 .

[6]  Yang Trista Cao,et al.  Toward Gender-Inclusive Coreference Resolution , 2019, ACL.

[7]  Michael S. Bernstein,et al.  Empath: Understanding Topic Signals in Large-Scale Text , 2016, CHI.

[8]  David Bamman,et al.  An Annotated Dataset of Coreference in English Literature , 2020, LREC.

[9]  A. Piper,et al.  Social Characters: The Hierarchy of Gender in Contemporary English-Language Fiction , 2019, Journal of Cultural Analytics.

[10]  Bronwyn M. Bjorkman.  Singular they and the syntactic representation of gender in English , 2017 .

[11]  Michael S. Bernstein,et al.  Shirtless and Dangerous: Quantifying Linguistic Signals of Gender Bias in an Online Fiction Writing Community , 2016, ICWSM.

[12]  Ryan Cotterell,et al.  Unsupervised Discovery of Gendered Language through Latent-Variable Modeling , 2019, ACL.

[13]  Matt Taddy,et al.  The Geometry of Culture: Analyzing the Meanings of Class through Word Embeddings , 2018, American Sociological Review.

[14]  Rachel Rudinger,et al.  Gender Bias in Coreference Resolution , 2018, NAACL.

[15]  Lauren Ackerman,et al.  Syntactic and cognitive issues in investigating gendered coreference , 2019 .

[16]  Dan Jurafsky,et al.  Content Analysis of Textbooks via Natural Language Processing: Findings on Gender, Race, and Ethnicity in Texas U.S. History Textbooks , 2020, AERA Open.

[17]  Yejin Choi,et al.  Connotation Frames of Power and Agency in Modern Films , 2017, EMNLP.

[18]  Rita Raley,et al.  Playing With Unicorns: AI Dungeon and Citizen NLP , 2020, Digit. Humanit. Q..

[19]  Benjamin S. Bloom,et al.  A Taxonomy for Learning, Teaching, and Assessing: A Revision of Bloom's Taxonomy of Educational Objectives , 2000 .

[20]  Jisun An,et al.  SemAxis: A Lightweight Framework to Characterize Domain-Specific Word Semantics Beyond Sentiment , 2018, ACL.

[21]  Nanyun Peng,et al.  The Woman Worked as a Babysitter: On Biases in Language Generation , 2019, EMNLP.

[22]  M. Long,et al.  Portrayals of Male and Female Scientists in Television Programs Popular Among Middle School-Age Children , 2010 .

[23]  Elias Benussi,et al.  How True is GPT-2? An Empirical Analysis of Intersectional Occupational Biases , 2021, ArXiv.

[24]  Jeffrey Dean,et al.  Efficient Estimation of Word Representations in Vector Space , 2013, ICLR.

[25]  Sameer Singh,et al.  Universal Adversarial Triggers for Attacking and Analyzing NLP , 2019, EMNLP.

[26]  David Bamman,et al.  A Bayesian Mixed Effects Model of Literary Character , 2014, ACL.

[27]  Arvind Narayanan,et al.  Semantics derived automatically from language corpora contain human-like biases , 2016, Science.

[28]  Katherine Elkins,et al.  Can GPT-3 Pass a Writer’s Turing Test? , 2020, Journal of Cultural Analytics.

[29]  Emily M. Bender,et al.  On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜 , 2021, FAccT.

[30]  Nanyun Peng,et al.  Towards Controllable Biases in Language Generation , 2020, FINDINGS.

[31]  M. Graham,et al.  Science faculty’s subtle gender biases favor male students , 2012, Proceedings of the National Academy of Sciences.

[32]  David Vandyke,et al.  Counter-fitting Word Vectors to Linguistic Constraints , 2016, NAACL.

[33]  Michael Mateas,et al.  Why Are We Like This?: The AI Architecture of a Co-Creative Storytelling Game , 2020, FDG.

[34]  David M. Mimno,et al.  Applications of Topic Models , 2017, Found. Trends Inf. Retr..

[35]  James Zou,et al.  Persistent Anti-Muslim Bias in Large Language Models , 2021, AIES.

[36]  Michael L. Littman,et al.  Measuring praise and criticism: Inference of semantic orientation from association , 2003, TOIS.

[37]  Brendan T. O'Connor,et al.  Learning Latent Personas of Film Characters , 2013, ACL.

[38]  Brian Larson,et al.  Gender as a Variable in Natural-Language Processing: Ethical Considerations , 2017, EthNLP@EACL.

[39]  Cameron Blevins,et al.  Jane, John ... Leslie? A Historical Method for Algorithmic Gender Prediction , 2015, Digit. Humanit. Q..

[40]  S. A. Inness.  Geek chic : smart women in popular culture , 2007 .

[41]  Brian A. Nosek,et al.  Harvesting implicit group attitudes and beliefs from a demonstration web site , 2002 .

[42]  Mark Chen,et al.  Language Models are Few-Shot Learners , 2020, NeurIPS.

[43]  Brendan T. O'Connor,et al.  Analyzing Gender Bias within Narrative Tropes , 2020, NLPCSS.