Masked Theme-Specific Named Entity Recognition Assisted with Large Language Models

Existing Named Entity Recognition (NER) methods are required to label relevant samples and train the concrete NER models. Due to the specification of theme-specific documents, these NER models are considerably hard to identify potential theme-specific entities. To address this challenge, we propose an effective two-stage approach of masked theme-specific NER associated with Large Language Models (LLMs), which uses the unsupervised mechanism rather than the supervised one. The approach involves theme-specific entity ontology construction and masked NER in heterogeneous documents. The first stage is associated with LLMs and Wikipedia category pages, and the second one is implemented with the masked NER based on the created ontology in the first stage. Extensive experimental results suggest that the proposed masked NER can precisely locate the known entities in the theme-specific entity ontology while improving the accuracy of NER in the remaining text. Compared to the mainstream NER frameworks such as spaCy 3, the masked NER can identify more valid entities in the input Markdown text and continuously use the newly detected unknown entities to update the created ontology.