Analyzing and comparing LM architectures for named entity recognition

Abstract

This project aims to compare and evaluate the performances of Encoder-only LLM, Decoder-only LLM and Encoder-Decoder LLM for Named Entity Recognition (NER). Despite Encoder-only models being known to be better than generative models such as Decoder-only models for NER tasks, due to the rapid growth of generative artificial intelligence development, there have been numerous research done to find ways to improve capabilities of these generative models in various fields, with NER being one of such fields. As such, this project aims to compare the difference in performance between the different model architectures as well as to seek a way to improve the performance of Encoder-Decoder/Decoder-only models to match that of Encoder-only models to potentially find a way to make NER tasks more accessible and cost-effective in practical applications.