Computing Semantic Similarity between Skill Statements for Approximate Matching

Proceedings of NAACL HLT 2007, pages 572–579, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics Computing Semantic Similarity between Skill Statements for Approxi- mate Matching Feng Pan Robert G. Farrell USC Information Sciences Institute IBM T. J. Watson Research Center Marina del Rey, CA 90292 Hawthorne, NY 10532 pan@isi.edu robfarr@us.ibm.com Abstract This paper explores the problem of com- puting text similarity between verb phrases describing skilled human behav- ior for the purpose of finding approximate matches. Four parsers are evaluated on a large corpus of skill statements extracted from an enterprise-wide expertise taxon- omy. A similarity measure utilizing com- mon semantic role features extracted from parse trees was found superior to an in- formation-theoretic measure of similarity and comparable to the level of human agreement. 1 Introduction Knowledge-intensive industries need to become more efficient at deploying the right expertise as quickly and smoothly as possible, thus it is desired to have systems that can quickly match and deploy skilled individuals to meet customer needs. The searches in most of the current matching systems are based on exact matches between skill state- ments. However, exact matching is very likely to miss individuals who are very good matches to the job but didn’t select the exact skills that appeared in the open job description. It is always hard for individuals to find the per- fect skills to describe their skill sets. For example, an individual might not know whether to choose a skill stating that refers to “maintaining” a given product or “supporting” it or whether to choose a skill about maintaining a “database” or about maintaining “DB2”. Thus, it is desirable for the job search system to be able to find approximate matches, instead of only exact matches, between available individuals and open job positions. More specifically, a skill similarity computation is needed to allow searches to be expanded to related skills, and return more potential matches. In this paper, we present our work on develop- ing a skill similarity computation based upon se- mantic commonalities between skill statements. Although there has been much work on text simi- larity metrics (Lin, 1998a; Corley and Mihalcea, 2005), most approaches treat texts as a bag of words and try to find shared words with certain statistical properties based on corpus frequencies. As a result, the structural information in the text is ignored in these approaches. We will describe a new semantic approach that takes the structural information of the text into consideration and matches skill statements on corresponding seman- tic roles. We will demonstrate that it can outper- form standard statistical text similarity techniques, and reach the level of human agreement. In Section 2, we first describe the skill state- ments we extracted from an enterprise-wide exper- tise taxonomy. In Section 3, we describe the performance of a standard statistical approach on this task. This motivates our semantic approach of matching skill statements on corresponding seman- tic roles. We also compare and evaluate the per- formance of four natural language parsers (the …

Content truncated to 500 words