Constructing query context knowledge bases for relevant term suggestion

Users are often not good at formulating queries that express their information needs. Since user queries are usually very short, it's helpful if relevant terms could be suggested. However, huge amount of query logs are usually needed to give useful suggestions. To improve query formulation and retrieval effectiveness in the absence of query logs, we propose to construct query context knowledge bases for relevant term suggestion based on pseudo relevance feedback from Web resources. Given a query, frequently co-occurring terms are first obtained from various search results such as blogs, news, and keyterms, and term candidates are extracted. Then, the contextual relevance of candidate terms is estimated by mutual information and Web n-gram language model. Finally, top-ranked terms of higher correlation with the query are selected. Experimental results show a high percentage of the most relevant terms can be suggested for single-word popular queries in top ranks.