A Progressive Learning Strategy for Medical Natural Language Understanding

ZHE YANG, Yi Huang, Mengfei Guo, Yaqin Chen, Xiaoting Wu, Junlan Feng, Chao Deng

Keywords: healthcare applications, clinical NLP, clinical knowledge

Abstract: Medical natural language understanding (NLU) tasks seek to extract clinically relevant information—such as diagnostic intent, symptomatic manifestations, laboratory findings, and therapeutic regimens—from medical dialogues or textual data. Regrettably, the paucity of annotated medical datasets often impedes the development of robustly trained models across diverse tasks. A promising approach involves decomposing neural networks into modular skill components, thereby facilitating the transfer of acquired knowledge from trained tasks to novel ones. Nevertheless, in multi-task learning frameworks, the indiscriminate aggregation of skill modules into a unified architecture may result in suboptimal skill refinement. To address this limitation, we introduce a progressive learning paradigm wherein each task is constrained to leverage only the network structures of tasks preceding it in a predefined difficulty hierarchy, thereby maximizing knowledge assimilation from less complex subtasks. For empirical validation, we select four pivotal medical NLP tasks: Single Sentence Intention Classification (SSIC), Sentence Pair Relationship Judgment (SPRJ), Named Entity Recognition (NER), and Classifying Positive and Negative Clinical Findings (CPNCF). Experimental results demonstrate that our proposed strategy yields consistent performance enhancements on the CPNCF task across multiple datasets.

Archival Status: Non-archival (not included in proceedings)

Submission Number: 48