【SME Onsite Academic Seminar】Relieving Racial Bias in Hate Speech Detection Datasets with a Small Number of Expert Annotations: A Prompt-based Learning Approach
Time and Date: 03:00 pm - 04:30 pm, November 17, 2023 (Friday)
Venue: Room 619, Teaching A Building
Speaker: Prof. Michael Chau (The University of Hong Kong)
Topic: Relieving Racial Bias in Hate Speech Detection Datasets with a Small Number of Expert Annotations: A Prompt-based Learning Approach
Meeting ID: 398 540 7949
About the Speaker
Dr. Michael Chau is a Professor in Innovation and Information Management of the HKU Business School at the University of Hong Kong. His research focuses on the cross-disciplinary intersection of information systems, computer science, business analytics, and information science, with an emphasis on the applications of data, text, and web mining in various business, education, and social domains. Dr. Chau’s research has been well published and recognized. His research has resulted in over 160 publications in high-quality journals and conferences and has received more than 8,000 citations. He is the recipient of multiple research achievement and best paper awards, and is highly ranked in several research productivity studies. He received a Ph.D. degree in management information systems from the University of Arizona and a B.Sc. degree in computer science and information systems from the University of Hong Kong. More information can be found on the web (http://www.business.hku.hk/~mchau/).
Hate speech is one of the major problems on social media platforms. Automatic hate speech detection methods relying on machine learning (ML) models, which learn from manually labeled datasets, have been proposed in both academia and industry. However, there is increasing evidence that hate speech detection datasets labeled by general annotators (e.g., amateurs or MTurk workers) contain systematic racial bias. When such biased datasets are used to train ML models, the resulting ML models will also be biased, possibly causing more harms to users. Unlike general annotators, experts have been found much less biased. However, expert annotations are time-consuming and expensive, and thus cannot be obtained efficiently. This paper bridges the gap by adopting a few-shot learning method based on using prompts on large language models (i.e., prompt-based learning). The proposed method utilizes a small number of expert annotations to debias a much larger dataset labeled by general annotators. Extensive experiments are conducted to demonstrate the superior debiasing performance of the proposed method using real-world data collected from social media platforms such as Twitter. The study has important academic and practical implications for hate speech detection and machine learning models.