摘要：

实体链接旨在将自然语言文本中的提及链接到知识库中相应的目标实体, 主要面临提及和候选实体的表征能力有限, 导致候选实体精确排序困难的问题, 而现有的知识库扩展和图嵌入等提高表征能力的方法依赖外部数据或知识, 限制了其应用。提出一种实体链接中提及和候选实体精确排序的方法, 通过结合提及上下文构建prompt问句, 将提及和候选实体相似度计算转化为基于prompt问句的打分模式。通过预训练模型微调打分器, 得到提及和候选实体相似度的打分, 并综合候选实体发现阶段的得分, 以筛选出更准确的目标实体。这一过程无需额外的知识, 能够融合上下文信息, 从而更准确地衡量提及和实体之间的相似度。在两个公共数据集上将该模型与基线模型进行实验比较, 结果表明, 相比次优模型, 该模型Acc@1值分别提升了0.88和0.41百分点。

关键词: 实体链接, prompt问句, 预训练模型, 实体消歧, 精确排序

Abstract:

Entity Linking(EL) aims to link mentions in natural language texts to corresponding target entities in the knowledge base. It mainly faces the problem of limited representation capabilities of mentions and candidate entities, which complicates the accurate ranking of candidate entities. Existing knowledge is based on expand methods, such as graph embedding, to improve representation capabilities by relying on external data or knowledge, which limits their applications. This study proposes a method for accurately sorting mentions and candidate entities in entity links, thereby constructing a prompt question by considering the mention context. The similarity calculation of mentions and candidate entities is converted into a scoring model based on the prompt question. The score is fine-tuned using the pretrained model, to obtain a similarity score between mentions and candidate entities. The scores in the candidate entity discovery phase are combined to filter out more accurate target entities. This process requires no additional knowledge and incorporates contextual information to accurately measure the similarity between mentions and entities. An experimental comparison has been conducted between the proposed and baseline models on two public datasets. The Acc@1 values of the proposed model has increased by 0.88 and 0.41 percentage points, respectively, with respect to those of the suboptimal model.

Key words: Entity Linking(EL), prompt question, pretrained model, entity disambiguation, precise rank