Zero-shot Implicit Fine-grained Named Entity Categorization with LLM-based Deep Pattern Mining

Named entity recognition (NER) mainly involves two phases: (i) detecting entities of interest and (ii) categorizing them. Named entity (NE) categorization, encompassing both coarse- and fine-grained categorization, has been evolving through various methods, including development focusing on effective data annotation and innovative learning approaches. However, NE categorization is still highly dependent on providing (i) a set of predefined categories to NER models and (ii) annotated data representing the target domain. Thus, the delivery of these limitations conflicts with broader automation, generalizability, and adaptability for the NER task. This paper introduces a framework that leverages a dual-processing approach to utilize the NLP advancements in LLMs and exploit latent spaces (LSs) innovatively, aiming to surpass these limitations. Precisely, this is achieved by (i) employing a deep patterns distillation technique that utilizes LLMs and (ii) conducting recursive fusion for these patterns' latent spaces using an encoder-decoder (LSED) architecture, thus resulting in facilitating the NER task with a zero-shot and implicit NE categorization framework. Further refinement procedures are mentioned as well.