Joint Textual Semantics and Contextual Attributes for Disambiguation of Named Entities in Power Marketing

With the development of the electric power industry and the deepening of market-oriented reform, electric power marketing audit plays a crucial role in ensuring the healthy operation of the market, and the audit work needs to mine potential problems and identify violations from a large amount of structured and unstructured data. However, traditional information retrieval and knowledge management tools face difficulties in dealing with complex business scenarios, especially in the named entity disambiguation task, where named entities in the text often have multiple meanings and different ex-pression forms, leading to entity confusion and information errors. Aiming at the above problems, this paper proposes a named entity disambiguation method based on the union of text semantics and background attributes. First, the contextual texts of entities to be disambiguated and their candidate entities are converted into word vector representations using Roberta model, and the text fusion similarity is generated by calculating cosine similarity and Jaccard similarity. Next, the attention mechanism is introduced to assign different weights to the contextual attributes of the entities to calculate the contextual similarity. Then, the text fusion similarity is weighted and fused with the background similarity to obtain the joint similarity score. Finally, the experimental results show that the proposed multivariate similarity fusion method in this paper achieves an accuracy of 92.43% on the power marketing audit dataset, provides more accurate entity disambiguation results, effectively improves the accuracy and robustness of entity disambiguation, and provides strong support for the construction of the knowledge graph of power marketing audit.