融合生成式模型的知识增强实体链指方法

未链接实体分类是实体链指(Entity Linking, EL)任务中的重要研究内容之一。现有方法存在上下文语义信息不充分、分类准确率低等问题，导致实体链指任务表现不佳。本研究提出一种融合生成式模型的知识增强实体链指方法。该方法将实体链指分为两个子模块，即候选实体排序模块和未链接实体分类模块。本研究基于高精度的候选实体排序模块，获得高质量的知识扩展信息，并对未链接实体分类任务进行知识增强；针对未链指实体提及的分类问题，提出一套生成式框架，该框架能够取得超过基线模型的性能。本研究方法在2020年全国知识图谱与语义计算大会(CCKS2020)评测任务二的中文短文本实体链指数据集上取得了目前最佳性能(整体F值为91.76%),证明知识增强和生成式框架的引入能提高模型的泛化能力，缓解未链接实体分类中的信息不充分问题。

Knowledge Enhanced Entity Linking Method Integrating Generative Model

Unlinked entity classification is one of the important research contents in Entity Linking (EL) task.The existing methods have problems such as insufficient contextual semantic information and low classification accuracy,which lead to poor performance of entity linking tasks.A knowledge-enhanced entity linking method integrating generative models is proposed in this study.This method divides the entity linking into two sub-modules,namely the candidate entity sorting module and the unlinked entity classification module.Based on the high-precision candidate entity ranking module,the high-quality knowledge expansion information is obtained and the knowledge of unlinked entity classification tasks is enhanced.Aiming at the classification problem mentioned by unlinked entities,a generative framework is proposed,which can achieve better performance than the baseline model.This research method has achieved the best performance on the Chinese short text entity linking dataset of China Conference on Knowledge Graph and Semantic Computing in 2020 (CCKS2020) evaluation task 2 (the overall F value is 91.76%),which proves that the introduction of knowledge enhancement and generative framework can improve the generalization ability of the model and alleviate the problem of insufficient information in unlinked entity classification.