식품 개체명 인식을 위한 트랜스포머 기반 모델과 대규모 언어 모델의 앙상블 기법

초록

Named Entity Recognition (NER) in the food domain is essential for accurate information extraction, but the low frequency of sparse entities makes training challenging. This study compares the performance of transformer-based models (BERT, ELECTRA) and large language models (LLMs, GPT-3.5-turbo, LLaMA 2 7B) to find the optimal combination. ELECTRA combined with CRF and bi-LSTM showed strong performance for sparse entities, while LLaMA 2 7B was effective in providing contextual support. ELECTRA+CRF achieved an F1-score of 0.89 for sparse entities, which improved to 0.91 when combined with LLaMA 2 7B using weighted voting. For general entities, the F1-score reached 0.93. The results demonstrate that combining transformer-based models with LLMs effectively enhances NER performance in the food domain.

키워드