Understanding the Semantic Intent of Natural Language Query

International Joint Conference on Natural Language Processing, pages 552–560, Nagoya, Japan, 14-18 October 2013. Understanding the Semantic Intent of Domain-Specific Natural Language Query Juan Xu, Qi Zhang, Xuanjing Huang Fudan University School of Computer Science {11210240066,qz,xjhuang}@fudan.edu.cn Abstract Queries asked on search engines nowa- days increasingly fall in full natural lan- guage, which refer to Natural Language queries (NL queries). Parsing that kind of queries for the purpose of understanding user’s query intent is an essential factor to search engine. To this end, a hierarchical structure is introduced to represent the se- mantic intent of NL query and then we fo- cus on the problem of mapping NL queries to the corresponding semantic intents. We propose a parsing method by conducting two steps as follows: (1) predicting se- mantic tags for a given input query; (2) building an intent representation for the query using the sequence of semantic tags based on a structured SVM classification model. Experimental results on a manu- ally labeled corpus show that our method achieved a sufficiently high result in term of precision and F1. 1 Introduction Nowadays, with the development of voice search, queries asked on search engines often fall in full natural language, which refer to NL queries. For example, for the purpose of looking for a restau- rant, it is natural for us to ask “find the best Ital- ian restaurant near seattle washington” rather than “Italian restaurant seattle wa”. This means that voice search users are liable to express their in- tent in natural language which differs significant- ly from Web users. In addition to the developing of voice search, the increasing use of smartphones with voice assistant further boosts the number of such natural language queries. This increasing amount of natural language queries brings a big challenge to search engines. Many search engines today, generally speaking, are based on matching keywords against struc- tured information from relational database. Con- sider the query “find the best Italian restaurant n- ear seattle washington”. Without understanding the semantic intent, it may retrieval some unsat- isfying results that merely contain all these key- words, which are not really search terms (e.g. restaurant). But when the search engine under- stands the intent of this query is to “find restauran- t”, and also knows the meanings of individual con- stituents (i.e the“restaurant” is head search terms, “the best Italian ” and “near seattle washington” are modifier), then it would be able to route the query to a specialized search module (in this case restaurant search) and return the most relevant and essential answers rather than results that merely contain all these keywords. In no small part, the success of such approach relies on robust understanding of query intent. Most previous works in this area focus on query tagging problem, i.e., assigning semantic labels to query terms (Li et al., 2009; Manshadi and Li, 2009; Sarkas et al., 2010; Bendersky et al., 2011). Indeed, with the label information, a search engine is able to provide users with more relevant …

Content truncated to 500 words