Gender and Dialect Bias in YouTube’s Automatic Captions

By Rachael Tatman. 2017

404 citations recorded by the source corpus.

This project evaluates the accuracy of YouTube’s automatically-generated captions across two genders and five dialect groups. Speakers’ dialect and gender was controlled for by using videos uploaded as part of the “accent tag challenge”, where speakers explicitly identify their language background. The results show robust differences in accuracy across both gender and dialect, with lower accuracy for 1) women and 2) speakers from Scotland. This finding builds on earlier research finding that speaker’s sociolinguistic identity may negatively impact their ability to use automatic speech recognition, and demonstrates the need for sociolinguistically-stratified validation of systems.

Subfields: Language and Linguistics, Natural Language Processing, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Language and Linguistics

In Natural Language Processing

In Safety, Robustness & Interpretability

Citations in the corpus

Gender Bias in Coreference Resolution (2018, methodology): “is explored by Beutel et al. (2017) and Zhang et al. (2018). Prior work also analyzes social and gender stereotyping in existing NLP and vision datasets (van Miltenburg, 2016; Rudinger et al., 2017). Tatman (2017) investigates the impact of gender and dialect on deployed speech recognition systems, ”
Designing Disaggregated Evaluations of AI Systems: Choices, Considerations, and Tradeoffs (2021, background): “However, several pieces of work have demonstrated that AI systems can perform differently for different groups of people, often exhibiting especially poor performance for already-disadvantaged groups (e.g., Nicol, Casey, and MacFarlane 2002; Rodger and Pendharkar 2004; Tatman 2017; Buolamwini and Ge”
MERLOT RESERVE: Neural Script Knowledge through Vision and Language and Sound (2022, methodology): “The automatic captions in YouTube are known to suffer from gender bias [107], which our model (like neural models generally) might in turn amplify [131].”
On Measuring Social Biases in Sentence Encoders (2019, background): “…gender in distributional semantic word representations (Her-belot et al., 2012), natural language inference data (Rudinger et al., 2017), and facial recognition systems (Buolamwini and Gebru, 2018), as well as at the intersection of dialect and gender in automatic speech recognition (Tatman, 2017).”
Gender Bias in Neural Natural Language Processing (2018, background): “uestion that we address in this paper. Prior work provides evidence of bias in autocomplete suggestions [Lapowsky, 2018] and differences in accuracy of speech recognition based on gender and dialect [Tatman, 2017] on popular online platforms. Word embeddings, initial pre-processors in many NLP tasks”
Racial Bias in Hate Speech and Abusive Language Detection Datasets (2019, background): “There are many examples of bias in supervised learning contexts: YouTube’s captioning models make more errors when transcribing women (Tatman, 2017), AAE is more likely to be misclassiﬁed as non-English by widely used language classiﬁers (Blodgett and O’Connor, 2017), numerous gender and racial…”
Mitigating Gender Bias in Natural Language Processing: Literature Review (2019, background): “Automatic speech detection works better with male voices than female voices (Tatman, 2017).”
Multi-Dimensional Gender Bias Classification (2020, background): “Negative gender stereotyping is known to be weakened or reinforced by the presence of other social factors, such as dialect (Tatman, 2017), class (Degaetano-”
Bias in Bios: A Case Study of Semantic Representation Bias in a High-Stakes Setting (2019, background): “ linking a gender pronoun to an occupation in which that gender is overrepresented than to an occupation in which it is underrepresented. Gender bias has also been studied in YouTube’s autocaptioning [38], where researchers found a higher word error rate for female speakers. In the context of langua”
Fairness Without Demographics in Repeated Loss Minimization (2018, background): “, 2017), and automatic video captioning (Tatman, 2017).”
WILDS: A Benchmark of in-the-Wild Distribution Shifts (2020, neutral): “These models can perform well on average but poorly on some users (Tatman, 2017; Caldas et al., 2018; Li et al., 2019b; Koenecke et al., 2020).”
Racial Disparity in Natural Language Processing: A Case Study of Social Media African-American English (2017, background): “tive to what men are saying. Access: search systems are necessary for people to access information online, and for videos they may depend on indexing text recognized from the audio. Tatman’s results [20] imply that such a search system will fail to •nd information produced by female speakers more o›”
Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization (2019, background): “However, models that are highly accurate on average can still consistently fail on rare and atypical examples (Hovy & Sgaard, 2015; Blodgett et al., 2016; Tatman, 2017; Hashimoto et al., 2018; Duchi et al., 2019).”
Rissanen Data Analysis: Examining Dataset Characteristics via Description Length (2021, background): “, 2000; Buolamwini and Gebru, 2018), at speech recognition for women and speakers from Scotland (Tatman, 2017), and at POS tagging for African American”
Conversational AI: Social and Ethical Considerations (2019, background): “…as bias is found in: detecting skin tones in pedestrians [52], predictive policing systems and justice [38], the display of STEM career ads [27], recidivism algorithms [2], politics of search engines [21], medical applications [13], automatic speech recognition [44], and in hiring algorithms [1].”
Five sources of bias in natural language processing (2021, background): “A much more widespread unintended negative consequence is the unfairness caused by demographic biases, such as unequal performance for different user groups (Tatman, 2017), misidentification of speakers and their needs (Criado Perez, 2019) or the proliferation of harmful stereotypes (e.”
An annotated dataset of literary entities (2019, methodology): “…as coreference resolution (Rudinger et al., 2018; Zhao et al., 2018), sentiment analysis (Kiritchenko and Mohammad, 2018), and speech recognition (Tatman, 2017), we can investigate the same phenomenon here: does a model trained on ACE result in a disparate impact in its performance recognizing…”