Named Entity Recognition in Chinese–English Speech Using Automatic Speech Recognition and Large Language Models
摘要
This research proposes an automated sensitive health information (SHI) entity identification system for Chinese and English code-switched clinical speech. The framework first applies to the WhisperX model for local automatic speech recognition to generate transcripts with word-level timestamps. SHI entities are then identified using the LLaMA3-70B model via an application programming interface. A two-stage inference design leverages tag chain-of-thought prompting to refine entity types and improve robustness on ambiguous mentions. Rule-based standardization and postprocessing include simplified-to-traditional conversion, spelling correction, duplicate truncation, regular-expression augmentation, and timestamp alignment, further enhancing output consistency. Experiments on the AI CUP 2025 task demonstrate that the framework achieves strong speech deidentification performance without task-specific model fine-tuning, offering a practical solution for privacy protection in bilingual medical data.