Medical short text classification via Soft Prompt-tuning

Abstract

In recent decades, medical short texts, such as medical conversations and online medical inquiries, have garnered significant attention and research. The advances in the medical short text have profound implications in practical applications, particularly for classifying in-patient discharge summaries and medical text reports, leading to improved understandability for medical professionals. However, the challenges posed by the short length, professional medical vocabulary, complex medical measures, and feature sparsity are further magnified in medical short text classification compared to general domains. This paper introduces a novel soft prompt-tuning method designed specifically for medical short text classification. Inspired by the recent success of prompt- tuning, which has been extensively explored to enhance semantic modeling in various natural language processing tasks with the appearance of GPT-3, our method incorporates an automatic template generation method to address the issues related to short length and feature sparsity. Additionally, we propose two different strategies to expand the label word space, effectively handling the challenges associated with specialized medical vocabulary and complex medical measures in medical short texts. The experimental results demonstrate the effectiveness of our method and its potential as a significant advancement in medical short text classification. By addressing issues related to short text length, feature sparsity, and specialized medical terminology, it offers a promising advancement toward more accurate and interpretable medical text classification.