Interest Analysis using PageRank and Social Interaction Content

International Joint Conference on Natural Language Processing, pages 912–916, Nagoya, Japan, 14-18 October 2013. Interest Analysis using PageRank and Social Interaction Content Chung-chi Huang Institute of Information Science Academia Sinica Taipei, Taiwan 115 u901571@gmail.com Lun-Wei Ku Institute of Information Science Academia Sinica Taipei, Taiwan 115 lwku@iis.sinica.edu.tw Abstract We introduce a method for learning to predict reader interest. In our approach, social inter- action content and both syntactic and seman- tic features of words are utilized. The pro- posed method involves estimating topical in- terest preferences and determining the informativity between articles and their social content. In interest prediction, we integrate articles’ quality social feedback representing readers’ opinions into articles to get infor- mation which may identify readers’ interests. In addition, semantic aware PageRank is used to find reader interest with the help of word interestingness scores. Evaluations show that PageRank benefits from proposed features and interest preferences inferred across arti- cles. Moreover, results conclude that social interaction content and the proposed selection process help to accurately cover more span of reader interest. 1 Introduction Web keyword extraction tools such as KEA (www.nzdl.org/Kea/) typically look at articles from authors’ perspective to calculate the im- portance of a word in articles. However, key- words are not necessarily words that interest readers. We found that articles could be analyzed more towards reader interest if a system exploit- ed social interaction content (e.g., reader feed- back) in social media. Consider the content of an example article. The Web post describes a newly-renovated old house and the history, life style, and surrounding sightseeing sites of a historical city where it is located. Most keyword tools can easily identify keywords the old house (謝宅) and the historical city (台南). However, article readers might also be interested in less frequent words like life style (生活) and traditional market (市場), and single- occurrence like rental fees (費用), which are also mentioned in most reader feedback. In the proposed method, an article was trans- formed into a word graph where vertices were words in the article and edges between vertices indicated words’ co-occurrences. To distinguish associate/key words from words of reader inter- est, readers’ quality interaction feedback was considered when building the word graph. Sub- sequently, word interest preferences and Pag- eRank were utilized to find interest terms. Weightings concerning syntactic and semantic features are utilized in PageRank. Moreover, content-source and content-word weighted Pag- eRank were exploited to return words for interest evaluation. The predicted interests can further be used as candidates for social tagging or article recommendation. 2 Related Work The state-of-the-art keyword extraction methods have been applied to a myriad of natural lan- guage processing tasks including document cate- gorization and summarization (Manning and Schutze, 2000; Litvak and Last, 2008), indexing (Li et al., 2004), information retrieval (Turney, 2000), and text mining on social networking or micro-blogging services (Li et al., 2010; Zhao et al., 2011;Wu et al., 2010). Here we extract key- words related to readers’ interests. Recently, collaborative tagging or social tag- ging has grown …

Content truncated to 500 words