Read this year's IJCAI article by Zhiyuan Liu, a teacher at Tsinghua University Representation learning for measuring Entity relatedness with Rich information.
Compared to the calculation of lexical similarity, the computation of the relevance of the wiki is more of a feature of the Wikipedia corpus. The correlation calculation is broadly divided into three categories:
1. Text-theoretic uses the massive nature of the Wikipedia corpus. Lexical characterization by means of statistical methods (Word representation). The traditional approach is simply to build a semantic space based on the wiki article (not knowing how to translate it properly). Statistical Terms and Wiki article, thus projecting words into a high-latitude semantic space. In recent years, we have been paying more attention to the expression of low-dimensional words through neural network language model.
Representative work: LSA ESA (Explicit Semantic analysis) Word2vec, etc.
2. graph-theoretic uses Wikipedia's link structure feature to map the Wikipedia entity. The intuitive understanding is that the closely connected entities in the diagram are more closely related.
3. information-theoretic uses Wikipedia's classification structure to structure and calculate the relevance of Wikipedia entities by their classification.
The method of calculating the similarity of words refers to three categories in the semantic relevance calculation based on the link structure and classification system of Chinese Wikipedia :
1. Based on large-scale corpora
2. Semantic dictionary based
3. Based on Wikipedia
Wiki-related calculation notes