Understanding Graph Structure of Wikipedia for Query Expansion

Abstract

Knowledge bases are very good sources for knowledge extraction, the ability to create knowledge from structured and unstructured sources and use it to improve automatic processes as query expansion. However, extracting knowledge from unstructured sources is still an open challenge. In this respect, understanding the structure of knowledge bases can provide significant benefits for the effectiveness of such purpose. In particular, Wikipedia has become a very popular knowledge base in the last years because it is a general encyclopedia that has a large amount of information and thus, covers a large amount of different topics. In this piece of work, we analyze how articles and categories of Wikipedia relate to each other and how these relationships can support a query expansion technique. In particular, we show that the structures in the form of dense cycles with a minimum amount of categories tend to identify the most relevant information

Query Expansion via structural motifs in Wikipedia Graph

“…In the expanded query 1, the triangular motif is used, in the expanded query 2 both the triangular and the square motifs are used to create the query graph and in the expanded query 3 only the square motif is used. We also compare our current results with those obtained in [11], which we use as an upper bound as they were achieved using a ground truth query graph (Upper bound Cycles length 3 and Upper bound Cycles length 4).…”

“…As a starting point for the work in this paper we take [11], where we use an information retrieval dataset, Image CLEF, to create a ground truth. For that purpose, for each Image CLEF request we extract the entities of its valid documents and match them with the articles in Wikipedia.…”

“…According to the analysis presented in [11], cycles of different lengths may favor better precision for small amounts of results, while other, for larger amount of results. Therefore, it is expected that, depending on the motif used to create the query graph, the titles of its articles serve better as expansion features for small or large tops of retrieved documents.…”

“…For these experiments, we focus on the Image CLEF dataset, since it is the dataset that was used in the analysis of [11], and that for which we have the ground truth. As baselines we use the input, which is the text issued by the user, the entities, whose input nodes have been manually selected and the input + entities.…”