Span Graph Based on Contrastive Learning for Nested Named Entity Recognition

To address the issue of nested entities in named entity recognition, this paper introduces a novel span classification model called SpCL. Unlike traditional models that solely consider the boundary information of a span, SpCL captures the internal features of the span by integrating features at multiple levels, including character-level, word-level, and context-level, in addition to the span’s width feature. Furthermore, SpCL captures external features of the span, especially its "similar entities", which have the same type or boundary words. SpCL uses contrastive learning to bring similar entities closer to each other, connects "similar entities" to construct a span graph, and employs a graph convolutional network to extract feature of the span sub-graph to enhance the current span representation. SpCL leverages the combined information of spans and corresponding "similar entities" to improve the performance of span classification model. Experimental results show that SpCL achieves a significant performance improvement, showing enhanced precision, recall, and F1 scores in comparison to other graph-based models and classical span classification models when applied to datasets such as ACE2004, ACE2005, and GENIA.