Query expansion is a method that suggests to users the keywords related to a query to help search activity of users. Since a query can have diverse meanings, query expansion has tried to cover the meanings by extracting information from retrieved documents. The previous studies have used clustering algorithms to group similar documents from retrieved documents, and have extracted a representative word as a keyword from each cluster. However, their main drawback is that they fixed the number of clusters regardless of queries even if each query can have various number of meanings.
This paper proposes a graph-based method to consider the characteristics of a query to suggest keywords. The main difference between the previous studies and our method is whether the number of meanings inherent in the query is fixed or not. The proposed method adopts a community detection algorithm which finds similar structures within a graph. Therefore we deal with a variety of meanings inherent in the query instead of fixing their number.
In this paper we regard a unit of structures as a word because it is easy to expand the keywords that are represented as words. For this, we first make a word co-occurrence graph from retrieved documents of a query. Second, we build word communities from the word co-occurrence graph using the community detection algorithm. Then, the representative word of each community is selected as a keyword. As a result, the keywords extracted from each community are used to expand the query.
In order to evaluate the proposed method, we compared our results to those of three baselines. The baselines are Google search engine, keyword recommendation using TF-IDF and ISKR. The evaluation results indicate that the proposed method outperforms the baselines with respect to relevance and diversity. This implies that the keywords derived from community detection algorithm are useful to expand queries.