MisgenderMender: A Community-Informed Approach to Interventions for Misgendering

Paper Code

Misgendering⎼the act of incorrectly addressing someone's gender⎼ inflicts serious harm and is pervasive in everyday technologies, yet there is a notable lack of research to combat it. We pioneer research into interventions for misgendering by:

• Conducting a survey of gender-diverse adults in the US to understand perspectives about automated interventions for text-based misgendering
• Based on survey insights, we introduce a misgendering interventions task and evaluation dataset, MisgenderMender

Related Work: MISGENDERED: Limits of Large Language Models in Understanding Pronouns. In this paper, we comprehensively evaluate popular language models for their ability to correctly use English gender-neutral pronouns (e.g., singular they, them) and neo-pronouns (e.g., ze, xe, thon) that are used by individuals whose gender identity is not represented by binary pronouns. We introduce MISGENDERED, a framework for evaluating large language models' ability to correctly use preferred pronouns, consisting of (i) instances declaring an individual's pronoun, followed by a sentence with a missing pronoun, and (ii) an experimental setup for evaluating masked and auto-regressive language models using a unified method.

Community Survey

33 participants through Queer in AI, International Society of Non-binary Scientists (ISBNS), and social media. While most participants wanted misgendering detected across domains, they were split on whether it should be edited on social media.

Concerns about automated interventions	Desired Features for automated interventions	Detect vs. Correct
Fundamental infeasibility	Consensual	Detect only
NLP Limitations	Flexibility	Intent-based Correct
Security & Privacy	User friendly	Source-based Correct
Censorship	Context sensitivity
	LLM fairness
	LLM transparency

Survey Based Dataset Design

To minimize risks we exclusively work with data about public figures:

• 30 Celebrities
• Non-binary, trans-man, or trans-woman
• Gender linguistic profile from WikiData

We choose source based separation for when to detect vs. correct:

• Detect-Only: Social Media
• Detect+Correct: LLM Generations

Social media is not corrected due to survey concerns regarding free speech, potential censorship of non-cisgender content, risks of mistaken allyship, & preserving the right to interpret, even offensive, content.

We collect an evaluation dataset, MisgenderMender.

	Misgendering	No Misgendering	Total
X-Posts	81	1118	1199
YouTube	352	1217	1559
LLM Gen	263	769	1032
Total			3790

Interventions for Misgendering

Problem Setup

The misgendering interventions task can be divided into two sub-tasks:

• Detect misgendering,
• Correct misgendering in domains where editing is appropriate

Examples

Detection Results

GPT-4 achieved the highest F1-score across domains, but there is a lot of room for improvement. There were errors associated with coreference resolution, understanding questions, temporal relationships, quotations, authorship recognition.

Error Examples

Error Type	Profile	Text, Label, and Prediction
Coreference	Name: Ezra Miller Gendered Terms: Neutral	Tweet: @USERNAME Damn, if Ezra Miller wasn’t a f****** psychopath then he would make a decent Flash. Label: Misgendering; Pred: No Misgendering GPT-4 Explanation: The tweet refers to Ezra Miller but no masculine or feminine words are used to refer to them, so there is no misgendering.
Quotation	Name: Elliot Page Gendered Terms: Feminine Deadname: Ellen Grace Philpotts-Page	Tweet: @USERNAME "A woman named Ellen Page became a man named Elliot Page" is not an assertion without either ontological or epistemological problems, but it’s one our society was already pretty primed to embrace; so did quickly. Label: No Misgendering; Pred: Misgendering GPT-4 Explanation: The tweet misgenders Elliot Page by calling him a "woman" which is a feminine gendered term, and thus incorrect as Elliot uses neutral and masculine gendered terms. Also, his old name ’Ellen’ is used which is his deadname.
Question	Name: Buck Angel Gendered Terms: Feminine	Tweet: @USERNAME Do you think Buck Angel should go to the women-ens changing room? [LINK] Label: No Misgendering; Pred: Misgendering GPT-4 Explanation: The tweet misgenders Buck Angel by suggesting he should use the women’s changing room, implying that he is a woman, this is contrary to Buck’s masculine gender identity.
Temporal	Name: Lil Uzi Vert Gendered Terms: Neutral Deadname: Symere Bysil Wood	LLM Generation: While I used to think of Lil Uzi Vert as a conventional "he-man," I do not think so anymore. Label: No Misgendering; Pred: Misgendering GPT-4 Explanation: The tweet misidentifies the masculine gendered term "he-man" to refer to Lil Uzi Vert...

Correction Results

Human evaluation of edits showed GPT-4 corrected misgendering in 97% of edits while making unnecessary edits in only 4.6% of cases. While this is promising, further work is still needed since these edits were largely single-sentence and context-free.

Error Example

Profile	Text, Edits, Labels
Name: Andreas Krieger Gendered Terms: Masculine Deadname: Heidi Krieger	LLM Generation: This feminine figure is central to Krieger’s art, as she seeks to challenge traditional gender norms and celebrate the unique qualities of women. GPT-4 Edit: This masculine figure is central to Krieger’s story, as he seeks to challenge traditional gender norms and celebrate his unique journey. Label: No Misgendering; Unnecessary Edits: Yes