Context-aware semantic classification of search queries for browsing community question-answering archives

Community Question-Answering (cQA) platforms have become massive repositories of user-generated content. To a great extent, these archives have proven to be highly re-usable. For instance, web search engines profit from their best answers for enhancing user experience when resolving question-like queries. Hence, considerable research efforts have gone into trying to revitalize and retrieve past answers contained in these archives. However, similarly to traditional web search, there is a linguistic gap between cQA questions and question-like search queries that are utilized for fetching information from these cQA repositories (e.g., “rib pain after ovulation” and “iron oxide household”). In fact, this gap does not only consider linguistic features, but also structural and social attributes. On the one hand side, cQA questions are long-winded, they can bear a title and a body, and community members are compelled to categorize questions at posting time. On the other hand side, search queries come as an uncategorized short stream of words. Moreover, in juxtaposition to cQA question, users typically submit streaks of semantically related search queries, when attempting to fulfil their information needs. This work digs deep into effectively exploiting semantic cues, yielded by preceding queries within the same user session, for classifying question-like search queries into twenty-six semantic cQA question categories. In order to find significant discriminative properties, we carried out experiments on a large-scale dataset acquired automatically. Broadly speaking, our results indicate that more effective semantic features can be computed as long as we account for a larger number of previous queries. In particular, facilitating Explicit Semantic Analysis for modelling the query context shows to be extremely helpful for increasing the classification rate.