关键词: 嵌套命名实体识别, 平面化句子表示, 多方向梯度, 特征提取, Sobel算子

Abstract: Abstract: Nested named entity recognition was regarded as a fundamental task in natural language processing. Span-based methods treated entity recognition as a span classification task, which effectively handled nested entities. Existing methods organized spans in a sentence into a two-dimensional plane, where each unit represented a span, similar to pixels in an image. Subsequently, edge detection techniques from image processing were integrated, and gradient operators were utilized to enhance and extract semantic edge features of entities in the planar sentence representation. However, existing gradient-based approaches overlooked multi-directional edge features between adjacent spans. To address the limitation, a nested named entity recognition method incorporating multi-directional gradient feature extraction was proposed. An eight-direction Sobel operator was employed, with entity positions treated as pixels in an image and the gradient properties of edges leveraged to extract more comprehensive and discriminative semantic edge features in the planar sentence representation. The proposed method achieves F1 scores of 88.01% and 81.23% on the ACE2005 Chinese dataset and the GENIA English dataset, respectively, demonstrating its effectiveness for nested named entity recognition tasks. Additionally, evaluation on the CoNLL2003 English flat dataset resulted in an F1 score of 92.52%, validating the scalability of the approach.

Key words: nested named entity recognition, planarized the sentence representation, multi directional gradient, feature extraction, Sobel operator