Nicholas Deas, Blake Vente, Amith Ananthram, Jessica A Grieser, Desmond U. Patton, Shana Kleiner, James R. Shepard Iii, Kathleen McKeown. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Warning: This paper contains content and language that may be considered offensive to some readers.While biases disadvantaging African American Language (AAL) have been uncovered in models for tasks such as speech recognition and toxicity detection, there has been little investigation of these biases for language generation models like ChatGPT.We evaluate how well LLMs understand AAL in comparison to White Mainstream English (WME), the encouraged "standard" form of English taught in American classrooms.We measure large language model performance on two tasks: a counterpart generation task, where a model generates AAL given WME and vice versa, as well as a masked span prediction (MSP) task, where models predict a phrase hidden from their input.Using a novel dataset of AAL texts from a variety of regions and contexts, we present evidence of dialectal bias for six pre-trained LLMs through performance gaps on these tasks.

Sharon Levy, Emily Allaway, Melanie Subbiah, Lydia Chilton, Desmond Patton, Kathleen McKeown, William Yang Wang. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

OBJECTIVE: The study provides considerations for generating a phenotype of child abuse and neglect in Emergency Departments (ED) using secondary data from electronic health records (EHR). Implications will be provided for racial bias reduction and the development of further decision support tools to assist in identifying child abuse and neglect. MATERIALS AND METHODS: We conducted a qualitative study using in-depth interviews with 20 pediatric clinicians working in a single pediatric ED to gain insights about generating an EHR-based phenotype to identify children at risk for abuse and neglect. RESULTS: Three central themes emerged from the interviews: (1) Challenges in diagnosing child abuse and neglect, (2) Health Discipline Differences in Documentation Styles in EHR, and (3) Identification of potential racial bias through documentation. DISCUSSION: Our findings highlight important considerations for generating a phenotype for child abuse and neglect using EHR data. First, information-related challenges include lack of proper previous visit history due to limited information exchanges and scattered documentation within EHRs. Second, there are differences in documentation styles by health disciplines, and clinicians tend to document abuse in different document types within EHRs. Finally, documentation can help identify potential racial bias in suspicion of child abuse and neglect by revealing potential discrepancies in quality of care, and in the language used to document abuse and neglect. CONCLUSIONS: Our findings highlight challenges in building an EHR-based risk phenotype for child abuse and neglect. Further research is needed to validate these findings and integrate them into creation of an EHR-based risk phenotype.

Child abuse and neglect are public health issues impacting communities throughout the United States. The broad adoption of electronic health records (EHR) in health care supports the development of machine learning-based models to help identify child abuse and neglect. Employing EHR data for child abuse and neglect detection raises several critical ethical considerations. This article applied a phenomenological approach to discuss and provide recommendations for key ethical issues related to machine learning-based risk models development and evaluation: (1) biases in the data; (2) clinical documentation system design issues; (3) lack of centralized evidence base for child abuse and neglect; (4) lack of "gold standard "in assessment and diagnosis of child abuse and neglect; (5) challenges in evaluation of risk prediction performance; (6) challenges in testing predictive models in practice; and (7) challenges in presentation of machine learning-based prediction to clinicians and patients. We provide recommended solutions to each of the 7 ethical challenges and identify several areas for further policy and research.

While natural language processing affords researchers an opportunity to automatically scan millions of social media posts, there is growing concern that automated computational tools lack the ability to understand context and nuance in human communication and language. This article introduces a critical systematic approach for extracting culture, context and nuance in social media data. The Contextual Analysis of Social Media (CASM) approach considers and critiques the gap between inadequacies in natural language processing tools and differences in geographic, cultural, and age-related variance of social media use and communication. CASM utilizes a team-based approach to analysis of social media data, explicitly informed by community expertise. We use of CASM to analyze Twitter posts from gang-involved youth in Chicago. We designed a set of experiments to evaluate the performance of a support vector machine using CASM hand-labeled posts against a distant model. We found that the CASM-informed hand-labeled data outperforms the baseline distant labels, indicating that the CASM labels capture additional dimensions of information that content-only methods lack. We then question whether this is helpful or harmful for gun violence prevention.

Ruiqi Zhong, Yanda Chen, Desmond Patton, Charlotte Selous, Kathy McKeown. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Serina Chang, Ruiqi Zhong, Ethan Adams, Fei-Tzin Lee, Siddharth Varia, Desmond Patton, William Frey, Chris Kedzie, Kathy McKeown. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018.

Violence is a serious problems for cities like Chicago and has been exacerbated by the use of social media by gang-involved youths for taunting rival gangs. We present a corpus of tweets from a young and powerful female gang member and her communicators, which we have annotated with discourse intention, using a deep read to understand how and what triggered conversations to escalate into aggression. We use this corpus to develop a part-of-speech tagger and phrase table for the variant of English that is used and a classifier for identifying tweets that express grieving and aggression.