Multi-source Autoregressive Entity Linking Based on Generative Method

The task of entity linking aims to map the entity reference in the text with the unambiguous entity in the knowledge base. However, under the background of the continuous growth of science and technology service platform services, the complexity and diversity of domain semantic information, and the vagueness and ambiguity of natural language, the task of linking industrial chain knowledge graphs and technology service resources faces the following problems: (1) The increase in the types of semantic information in the do-main will lead to the lack of cognition and ambiguity in the entity linking process, which will affect the accuracy of entity linking; (2) the context and interaction of corpus information Dependencies become more complex, making insufficient consideration of the mapping relationship between linked entities. To address the above issues, this paper proposes a generative method and multiple sources of information Autoregressive Entity Linking (GMoAEL), which uses data sets such as unstructured textual descriptions, related reference links, and fine-grained structured types Information and other sources of information are used to build a unified dense representation for entity learning, and an autoregressive model in the form of encoder-decoder is used to adjust the link generation process to handle the complex mapping relationship between in-put and output. This paper uses the AIDA CoNLL-YAGO data set to conduct experiments. Compared with other methods, the Micro-F1 value is 1.6% points ahead, which verifies the feasibility and effectiveness of this method.