Beyond GPT-NER: ChatGPT as Ensemble Arbitrator for Discontinuous Named Entity Recognition in Health Corpora

In medicine and healthcare, NER (Named Entity Recognition) involves identifying clinically relevant entities such as medications, symptoms, and adverse drug events (ADEs). This task is particularly challenging due to discontinuous NER (DNER), fragmented entities that may span multiple sentences. To date, methods to address DNER have not been explored using ensemble learning to the best of our knowledge. The advent of large language models (LLMs) like ChatGPT has demonstrated significant potential across various NLP tasks. However, most studies use ChatGPT as a problem-solving tool, such as the well-known GPT-NER, rather than integrating it into ensemble learning. This paper proposes a novel ensemble method incorporating ChatGPT as an arbitrator to ensemble five state-of-the-art (SOTA) NER models through custom prompt engineering. An experiment conducted over three benchmark medical datasets, CADEC, ShARe13, and ShARe14, indicates that our proposed ChatGPT-coordinate ensemble outperformed SOTA models, achieving F1-score improvements of 1.13%, 0.54%, and 0.67% on the CADEC, ShARe13, and ShARe14 datasets, respectively. Compared to the voting ensemble, our method improved scores by 0.63%, 0.32%, and 0.09%, while achieving average F1-score gains of 7.42%, 0.89%, and 0.54% over GPT-3.5 and GPT-4.