Wikipedia entity retrieval with word and entity embedding space alignment

Abstract

User queries used to search on Wikipedia might not always correspond to a Wikipedia article. Searching for such queries on the built-in search engine for Wikipedia does not return as many relevant Wikipedia articles as expected, since it is only text-based, and it overlooks the relationship between different Wikipedia articles. To improve the results of the Wikipedia search engine, with the help of the Wikipedia Link Graph, we propose a simple word – entity embedding space alignment model for searching relevant Wikipedia articles using fringe keywords from the Computer Science domain. By using this model, our problem can be reduced to finding the closest neighbors to a user query translated from the word space to the entity space. Due to a lack of dataset available for this purpose, a custom dataset constructed from Wikipedia2Vec is used to evaluate the model. Empirically, our word – entity alignment model does not always fair better than the text-based models.