Large language models have demonstrated outstanding capabilities in handling various tasks by interpreting and responding to prompt instructions. In this paper, we delve into the application of instruction-based fine-tuning of large language models for the task of Named Entity Recognition (NER). We propose an approach to explore virtual label words for entity categories, where the pre-trained language model automatically explores words related to labels and then applies these more easily understood virtual labels to the NER task, thus optimizing the model's accuracy in identifying specific entities. Specifically, we first engage in label engineering, using a pre-trained language model to establish connections between the initial label words and a series of relevant entity words based on the output distribution of the pre-trained language model. Virtual label words are constructed and filtered to enhance the semantic representation of the label words. The NER task is then reconfigured into a text2text format to better utilize the knowledge within the pre-trained language model using a unified information extraction framework, leveraging the previously obtained label prototypes (or virtual labels) for the NER task. This method enhances the semantic representation of label words by establishing connections between the initial label words and a series of related entity words within the original unified information extraction framework and prediction paradigm. To validate the performance of the proposed model, we conduct experiments on multiple datasets, and the results demonstrate that this method effectively enriches the model's semantic understanding capabilities.
Paper
Full text
Named Entity Recognition via Unified Information Extraction Framework
Semantic Scholar · 2024
Abstract
Large language models have demonstrated outstanding capabilities in handling various tasks by interpreting and responding to prompt instructions. In this paper, we delve into the application of instruction-based fine-tuning of large language models for the task of Named Entity Recognition (NER). We propose an approach to explore virtual label words for entity categories, where the pre-trained language model automatically explores words related to labels and then applies these more easily understood virtual labels to the NER task, thus optimizing the model's accuracy in identifying specific entities. Specifically, we first engage in label engineering, using a pre-trained language model to establish connections between the initial label words and a series of relevant entity words based on the output distribution of the pre-trained language model. Virtual label words are constructed and filtered to enhance the semantic representation of the label words. The NER task is then reconfigured into a text2text format to better utilize the knowledge within the pre-trained language model using a unified information extraction framework, leveraging the previously obtained label prototypes (or virtual labels) for the NER task. This method enhances the semantic representation of label words by establishing connections between the initial label words and a series of related entity words within the original unified information extraction framework and prediction paradigm. To validate the performance of the proposed model, we conduct experiments on multiple datasets, and the results demonstrate that this method effectively enriches the model's semantic understanding capabilities.