A Generative Large Language Model–based Approach for Sensitive Data Identification in Medical Speech

This paper presents the methodology and results of our entry (TEAM_7897) in the AI CUP 2025 Spring Season “Medical Speech Sensitive Personal Data Identification” competition. The primary challenge was to accurately extract sensitive personal information and corresponding timestamps from medical dialogue transcripts. To address this, we propose a comprehensive pipeline that reframes the named entity recognition task as a generative problem using a large language model. At the core of the system is the Qwen3-8B model, fine-tuned efficiently via low-rank adaptation to recognize and format entities. A key innovation is a two-step data augmentation strategy designed to mitigate the severe class imbalance in the dataset: first generating novel entity examples, then producing realistic contextual sentences. Automatic speech recognition was performed using openai/whisper-large-v3. Notably, the native Whisper model (mixed error rate: 0.152) outperformed its fine-tuned version, and the NER model trained solely on the original dataset surpassed the model trained with augmented data. Analysis revealed that limitations in the quality and diversity of the generated data hindered performance. These findings demonstrate the potential of generative LLMs for complex NER tasks while emphasizing that data quality is critical for effective augmentation pipelines.