An official website of the United States government
We investigate how annotators’ insensitivity to differences in dialect can lead to racial bias in automatic hate speech detection models, potentially amplifying harm against minority populations. We first uncover unexpected correlations between surface markers of African American English (AAE) and ratings of toxicity in several widely used hate speech datasets. Then, we show that models trained on these corpora acquire and propagate these biases, such that AAE tweets and tweets by self-identified African Americans are up to two times more likely to be labelled as offensive compared to others. Finally, we propose dialect and race priming as ways to reduce the racial bias in annotation, showing that when annotators are made explicitly aware of an AAE tweet’s dialect they are significantly less likely to label the tweet as offensive.
more »
« less
This content will become publicly available on June 7, 2026
Large Language Model Annotation Bias in Hate Speech Detection
Large language models (LLMs) are fast becoming ubiquitous and have shown impressive performance in various natural language processing (NLP) tasks. Annotating data for downstream applications is a resource-intensive task in NLP. Recently, the use of LLMs as a cost-effective data annotator for annotating data used to train other models or as an assistive tool has been explored. Yet, little is known regarding the societal implications of using LLMs for data annotation. In this work, focusing on hate speech detection, we investigate how using LLMs such as GPT-4 and Llama-3 for hate speech detection can lead to different performances for different text dialects and racial bias in online hate detection classifiers. We used LLMs to predict hate speech in seven hate speech datasets and trained classifiers on the LLM annotations of each dataset. Using tweets written in African-American English (AAE) and Standard American English (SAE), we show that classifiers trained on LLM annotations assign tweets written in AAE to negative classes (e.g., hate, offensive, abuse, racism, etc.) at a higher rate than tweets written in SAE and that the classifiers have a higher false positive rate towards AAE tweets. We explore the effect of incorporating dialect priming in the prompting techniques used in prediction, showing that introducing dialect increases the rate at which AAE tweets are assigned to negative classes.
more »
« less
- Award ID(s):
- 2228616
- PAR ID:
- 10612700
- Publisher / Repository:
- Proceedings of the International AAAI Conference on Web and Social Media
- Date Published:
- Journal Name:
- Proceedings of the International AAAI Conference on Web and Social Media
- Volume:
- 19
- ISSN:
- 2162-3449
- Page Range / eLocation ID:
- 1389 to 1418
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Hate speech and offensive language are rampant on social media. Machine learning has provided a way to moderate foul language at scale. However, much of the current research focuses on overall performance. Models may perform poorly on text written in a minority dialectal language. For instance, a hate speech classifier may produce more false positives on tweets written in African-American Vernacular English (AAVE). To measure these problems, we need text written in both AAVE and Standard American English (SAE). Unfortunately, it is challenging to curate data for all linguistic styles in a timely manner—especially when we are constrained to specific problems, social media platforms, or by limited resources. In this paper, we answer the question, “How can we evaluate the performance of classifiers across minority dialectal languages when they are not present within a particular dataset?” Specifically, we propose an automated fairness fuzzing tool called FuzzE to quantify the fairness of text classifiers applied to AAVE text using a dataset that only contains text written in SAE. Overall, we find that the fairness estimates returned by our technique moderately correlates with the use of real ground-truth AAVE text. Warning: Offensive language is displayed in this manuscript.more » « less
-
Existing large language models (LLMs) that mainly focus on Standard American English (SAE) often lead to significantly worse performance when being applied to other English dialects. While existing mitigations tackle discrepancies for individual target dialects, they assume access to high-accuracy dialect identification systems. The boundaries between dialects are inherently flexible, making it difficult to categorize language into discrete predefined categories. In this paper, we propose DADA (Dialect Adaptation via Dynamic Aggregation), a modular approach to imbue SAE-trained models with multi-dialectal robustness by composing adapters which handle specific linguistic features. The compositional architecture of DADA allows for both targeted adaptation to specific dialect variants and simultaneous adaptation to various dialects. We show that DADA is effective for both single task and instruction finetuned language models, offering an extensible and interpretable framework for adapting existing LLMs to different English dialects.more » « less
-
null (Ed.)Abstract Objective Machine learning is used to understand and track influenza-related content on social media. Because these systems are used at scale, they have the potential to adversely impact the people they are built to help. In this study, we explore the biases of different machine learning methods for the specific task of detecting influenza-related content. We compare the performance of each model on tweets written in Standard American English (SAE) vs African American English (AAE). Materials and Methods Two influenza-related datasets are used to train 3 text classification models (support vector machine, convolutional neural network, bidirectional long short-term memory) with different feature sets. The datasets match real-world scenarios in which there is a large imbalance between SAE and AAE examples. The number of AAE examples for each class ranges from 2% to 5% in both datasets. We also evaluate each model's performance using a balanced dataset via undersampling. Results We find that all of the tested machine learning methods are biased on both datasets. The difference in false positive rates between SAE and AAE examples ranges from 0.01 to 0.35. The difference in the false negative rates ranges from 0.01 to 0.23. We also find that the neural network methods generally has more unfair results than the linear support vector machine on the chosen datasets. Conclusions The models that result in the most unfair predictions may vary from dataset to dataset. Practitioners should be aware of the potential harms related to applying machine learning to health-related social media data. At a minimum, we recommend evaluating fairness along with traditional evaluation metrics.more » « less
-
ISCA (Ed.)In this paper, we explore automatic prediction of dialect density of the African American English (AAE) dialect, where dialect density is defined as the percentage of words in an utterance that contain characteristics of the non-standard dialect. We investigate several acoustic and language modeling features, including the commonly used X-vector representation and ComParE feature set, in addition to information extracted from ASR transcripts of the audio files and prosodic information. To address issues of limited labeled data, we use a weakly supervised model to project prosodic and X-vector features into low-dimensional task-relevant representations. An XGBoost model is then used to predict the speaker's dialect density from these features and show which are most significant during inference. We evaluate the utility of these features both alone and in combination for the given task. This work, which does not rely on hand-labeled transcripts, is performed on audio segments from the CORAAL database. We show a significant correlation between our predicted and ground truth dialect density measures for AAE speech in this database and propose this work as a tool for explaining and mitigating bias in speech technology.more » « less
