Improving Zero-Shot Multiclass Classification for Narrative Reports from National Violent Death Reporting System

Authors

Yijun Shao, George Washington University, Washington, DC, USA.
Ryan Wu, George Washington University, Washington, DC, USA.
Adnan Lakdawala, George Washington University, Washington, DC, USA.
Alicia M. Jones, Northwestern University, Chicago, IL, USA.
Megan Koch, Illinois Department of Public Health, Springfield, IL, USA.
Yingxuan Liu, Northwestern University, Chicago, IL, USA.
Lori A. Post, Northwestern University, Chicago, IL, USA.
Maryann Mason, Northwestern University, Chicago, IL, USA.
Qing Zeng-Treitler, George Washington University, Washington, DC, USA.

Document Type

Journal Article

Publication Date

8-7-2025

Journal

Studies in health technology and informatics

Volume

329

DOI

10.3233/SHTI250939

Keywords

Zero-shot learning; large language model; text classification

Abstract

The natural language processing pipeline powered by the BART model is a popular zero-shot text classification system. While the standard approach for using this pipeline can achieve impressive accuracies in many multiclass classification tasks, we believed that there was still room to improve and developed an improved approach for that. Both approaches were used for the classification of narrative reports, and the results showed that the improved approach could increase the accuracies significantly over the standard approach. The improved approach is made general and can be applied to other use cases as well.

APA Citation

Shao, Yijun; Wu, Ryan; Lakdawala, Adnan; Jones, Alicia M.; Koch, Megan; Liu, Yingxuan; Post, Lori A.; Mason, Maryann; and Zeng-Treitler, Qing, "Improving Zero-Shot Multiclass Classification for Narrative Reports from National Violent Death Reporting System" (2025). GW Authored Works. Paper 7805.
https://hsrc.himmelfarb.gwu.edu/gwhpubs/7805

Department

Clinical Research and Leadership