Computing text semantic relatedness using the contents and links of a hypertext encyclopedia

We propose a method for computing semantic relatedness between words or texts by using knowledge from hypertext encyclopedias such as Wikipedia. A network of concepts is built by filtering the encyclopedia’s articles, each concept corresponding to an article. Two types of weighted links between concepts are considered: one based on hyperlinks between the texts of the articles, and another one based on the lexical similarity between them. We propose and implement an efficient random walk algorithm that computes the distance between nodes, and then between sets of nodes, using the visiting probability from one (set of) node(s) to another. Moreover, to make the algorithm tractable, we propose and validate empirically two truncation methods, and then use an embedding space to learn an approximation of visiting probability. To evaluate the proposed distance, we apply our method to four important tasks in natural language processing: word similarity, document similarity, document clustering and classification, and ranking in information retrieval. The performance of the method is state-of-the-art or close to it for each task, thus demonstrating the generality of the knowledge resource. Moreover, using both hyperlinks and lexical similarity links improves the scores with respect to a method using only one of them, because hyperlinks bring additional real-world knowledge not captured by lexical similarity.