Multi-Span Matrices and Knowledge Distillation for Nested Named Entity Recognition

Named Entity Recognition (NER) tasks aim to identify and classify named entities in text. When entities overlap, the task is referred to as nested named entity recognition (nested NER). Span-based methods are commonly used for nested NER. This method enumerates all entity spans in the text, identifying and classifying all possible entity types. However, previous work has overlooked the problem of prediction bias caused by the imbalance of entity category labels, and the problem of excessive non-entities leading to insufficient entity information. To address the aforementioned problem, we propose a nested NER method based on multi-span matrices and knowledge distillation. First, we treat entity categories as basic units and generates multiple span matrices. Each span matrix responsible for recognizing one type of entity category, thereby addressing the imbalance in entity category labels. Second, to resolve the problem of excessive non-entities resulting in sparse entity information in the multispan matrices, we introduce a knowledge distillation method. Through knowledge distillation, it injects entity information from single-span matrices into multi-span matrices, enhancing the entity information in the latter. To validate the effectiveness of the model, we conducted tests on multiple NER datasets. The experimental results show that the proposed method performs better in terms of entity recognition accuracy compared to related work.