A Study On Semantic Tagging Of Chinese Product Query Based On Conditional Random Fields

Users usually use shopping search engine to find products they like, to compareproduct information in search results, such as price, sales and buyers’ reviews. Whenclick one product item on the listing, they will get more detail information on productdetail pages. Understanding the intents of these product queries can not only improvea user’s search experience, but also boost a site’s advertising profits. As one steptoward this goal, we study the problem of semantic tagging of product query, which isto assign each query segment/word to a pre-defined semantic category. Usingstatistical sequence labeling models to tag product queries has been shown to performwell. For instance, Conditional Random Fields (CRFs) model has achievedstate-of-the-art performance on such research topic.Product query has the following three aspects: short and lack of features; productqueries of different categories are significantly different; the words in product queryare context-sensitive and can be conveniently divided into semantic categories.Different from many other studies on semantic tagging of product query, this paperfocuses on Chinese-based product query, but not English-based. The main workcontains three parts. Firstly, many types of feature functions and feature templateshave been combined to improve semantic tagging performance of Chinese productquery. Secondly, this paper proposes a new approach to conduct semantic lexicons byusing product titles from in-domain database, which will further improve performance,especially for the lack of training samples. Last but not the least, as we known, Averaged Perceptron (AP) algorithm has been applied to training the semantic taggingmodels for the first time, which can not only reach better performance, but alsoshorten training time dramatically. This finding will make this study more practical.Furthermore, this paper conducts various experiments to compare taggingperformance under different situations, such as different models, tagging withsegmented queries or un-segmented queries, different feature templates, differentparameter estimate methods, and using lexicon-based features or not. Experimentresults and analysis provide many valuable benchmarks for semantic tagging ofChinese product query.