Unsupervised Template Mining for Semantic Category Understanding

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 799–809, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Unsupervised Template Mining for Semantic Category Understanding Lei Shi1,2∗, Shuming Shi3, Chin-Yew Lin3, Yi-Dong Shen1, Yong Rui3 1State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences 2University of Chinese Academy of Sciences 3Microsoft Research {shilei,ydshen}@ios.ac.cn {shumings,cyl,yongrui}@microsoft.com Abstract We propose an unsupervised approach to constructing templates from a large collec- tion of semantic category names, and use the templates as the semantic representa- tion of categories. The main challenge is that many terms have multiple meanings, resulting in a lot of wrong templates. Sta- tistical data and semantic knowledge are extracted from a web corpus to improve template generation. A nonlinear scoring function is proposed and demonstrated to be effective. Experiments show that our approach achieves significantly better re- sults than baseline methods. As an imme- diate application, we apply the extracted templates to the cleaning of a category col- lection and see promising results (preci- sion improved from 81% to 89%). 1 Introduction A semantic category is a collection of items shar- ing common semantic properties. For example, all cities in Germany form a semantic category named “city in Germany” or “German city”. In Wikipedia, the category names of an entity are manually edited and displayed at the end of the page for the entity. There have been quite a lot of approaches (Hearst, 1992; Pantel and Ravichan- dran, 2004; Van Durme and Pasca, 2008; Zhang et al., 2011) in the literature to automatically extract- ing category names and instances (also called is-a or hypernymy relations) from the web. Most existing work simply treats a category name as a text string containing one or multiple words, without caring about its internal structure. In this paper, we explore the semantic structure of category names (or simply called “categories”). ∗This work was performed when the first author was vis- iting Microsoft Research Asia. For example, both “CEO of General Motors” and “CEO of Yahoo” have structure “CEO of [com- pany]”. We call such a structure a category tem- plate. Taking a large collection of open-domain categories as input, we construct a list of category templates and build a mapping from categories to templates. Figure 1 shows some example semantic categories and their corresponding templates. Templates can be treated as additional features of semantic categories. The new features can be exploited to improve some upper-layer applica- tions like web search and question answering. In addition, by linking categories to templates, it is possible (for a computer program) to infer the se- mantic meaning of the categories. For example in Figure 1, from the two templates linking to cat- egory “symptom of insulin deficiency”, it is rea- sonable to interpret the category as: “a symptom of a medical condition called insulin deficiency which is about the deficiency of one type of hor- mone called insulin.” In this way, our knowledge about a category can go …

Content truncated to 500 words