Topical Key Concept Extraction from Folksonomy

International Joint Conference on Natural Language Processing, pages 480–488, Nagoya, Japan, 14-18 October 2013. Topical Key Concept Extraction from Folksonomy Han Xue1,2, Bing Qin1*, Ting Liu1, Chao Xiang1 1Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, Harbin, China {hxue,bqin,tliu}@ir.hit.edu.cn, Cloudaice@gmail.com 2Harbin Engineering University, Harbin, China Abstract  Concept extraction is a primary subtask of on- tology construction. It is difficult to extract new concepts from traditional text corpus. Moreover, building a single ontology for mul- tiple-topic corpus may lead to misconception. To deal with these problems, this paper pro- poses a novel framework to extract topical key concepts from folksonomy. Folksonomy is a valuable data source due to real-time update and rich user-generated contents. We first identify topics from folksonomy using topic models. Next the tags are ranked according to their importance for a certain topic by apply- ing topic-specific random walk methods. The top-ranking tags are extracted as topical key concepts. Especially, a novel link weight func- tion which combines the local structure infor- mation and global semantic similarity is pro- posed in importance score propagation. From the perspectives of qualitative and quantitative investigation, our method is feasible and effec- tive. 1 Introduction Ontology can be seen as an organized structure of concepts according to their relations (Cui et al., 2009). Therefore, concept extraction is an im- portant subtask of ontology construction. Exist- ing works mainly focus on extracting concepts from text corpus (Buitelaar et al., 2005). Howev- er, it is difficult to find text corpus that accurate- ly characterize a highly focused, even fast- changing topic (Liu et al., 2012) of the domain. For instance, it is easier to find text corpus for a common topic of movie such as “comedy”, but it  *Correspondence author is much more difficult to find one for a specific topic such as “cult”. Since “cult” movies often do not follow traditional standards of mainstream movies. Moreover, it is not easy for them to find a formal definition and description in text corpus. However, we can more easily find some tags (ar- bitrary words assigned by people to the resources of interest) to describe this kind of movie, such as cult, non-mainstream, small budgets and so on. Motivated by the fact that social tags give us flexibility and ease to describe a topic, we try to use folksonomy (Trant, 2009) as a new data source. The word ‘folksonomy’ is a blend of the words ‘folk’ and ‘taxonomy’. It is the achieve- ment of collective wisdom derived from the practice of collaboratively creating tags to anno- tate and categorize web resource. Figure 1. A folksonomy example Take Douban.com for example, which is a Chinese SNS website allowing registered users to record information and create tags related to Website Resource Tag set 480 their interested resources, such as film, books, music and recent activities. As shown in Fig. 1, in the folksonomy-driven web site 豆瓣电影网 站‘Douban.com Movie1’, the resource 致我们终 将逝去的青春‘So Young’ is annotated with a set of tags including …

Content truncated to 500 words