Enhancing Named Entity Recognition with a BERT-DQN Hybrid Model

Abstract

Named Entity Recognition (NER) is a fundamental task in natural language processing (NLP) that involves identifying and classifying proper nouns into predefined categories such as names of people, organizations, locations, etc. This paper proposes an advanced NER model by integrating Bidirectional Encoder Representations from Transformers (BERT) with Conditional Random Fields (CRF), further enhanced with Deep Q-Network (DQN) reinforcement learning within the CRF layer. Utilizing the CoNLL 2003 dataset, our model aims to leverage the powerful feature extraction capabilities of BERT along with the structured prediction advantage of CRF. The DQN integration aims to optimize the sequence tagging decisions dynamically, addressing limitations found in traditional NER models. Our results demonstrate significant improvements in precision, recall, and F1 score, highlighting the efficacy of incorporating reinforcement learning into the BERT-CRF architecture for NER tasks.