Query Intent Detection By Mining Click-Through Data

Search engine has been an indispensable tool when people seek information on the Internet. Along with the development of society and improvement of technology, search results become better and better, from literally matching their queries to semantically matching. Queries are the most important information that users give to search engine. Through query intent detection, search engine can return the very results user wants and improve the user experience, even though the queries are short.Based on CIKM 2014 Query Intent Detection Competition, this paper introduce the significance and challenges of the task and resolve the problem of shortness of queries and lack of enough information through mining the click-through data. Another challenge is that the data has been encoded and common word segmentation algorithm can't been easily used, so the bag-of-words model is built with k-gram and end-gram and information gain. Then the feature vector is calculated with tfidf for document representation. Finally, linear support vector machine is used to classify queries. There are 7 predefined categories in this task and one query can be labelled with 2 categories, so this paper train 7 binary classifier with One-Vs-Rest strategy and predict the categories through the probabilities.Lots of experiments are conducted to evaluate the parameters of procedures and we use five-fold cross validation to tune the parameters of SVM. This method is simple, effective and efficient. We place six among 530 teams with F1 score 0.9181 on the final leaderboard.