Keyphrase Extraction from Online News Using Binary Integer Programming

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 165–173, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Keyphrase Extraction from Online News Using Binary Integer Programming Zhuoye Ding, Qi Zhang, Xuanjing Huang Fudan University School of Computer Science {09110240024,qz,xjhuang}@fudan.edu.cn Abstract In recent years, keyphrase extraction has received great attention, and been success- fully employed by various applications. Keyphrases extracted from news articles can be used to concisely represent main contents of news events. Keyphrases can help users to speed up browsing and find the desired contents more quickly. In this paper, we first present several crite- ria of high-quality news keyphrases. Af- ter that, in order to integrate those crite- ria into the keyphrase extraction task, we propose a novel formulation which con- verts the task to a binary integer program- ming problem. The formulation cannot only encode the prior knowledge as con- straints, but also learn constraints from data. We evaluate the proposed approach on a manually labeled corpus. Experi- mental results demonstrate that our ap- proach achieves better performances com- pared with the state-of-the-art methods. 1 Introduction Keyphrase extraction is a long studied topic in natural language processing. A keyphrase, which consists a word or a group of words, is defined as a precise and concise expression of one or more documents. It has been widely used in var- ious applications such as summarization, cluster- ing, categorizing, browsing, and so on. In recent years, keyphrase extraction has received much at- tention (Witten et al., 1999; Zha, 2002; Hulth, 2003; Tomokiyo and Hurst, 2003; Chen et al., 2005; Medelyan et al., 2009; Liu et al., 2009). Keyphrases are usually manually chosen by authors, for scientific publications, magazine ar- ticles, books, et al. Due to the expensive and time consuming effort of manually assign- ing keyphrase, web pages and online news rarely contain keyphrases. It should be useful to au- tomatically extract keyphrases from online news to represent their main contents. There are al- ready a number of studies which focus on extract- ing keyphrases from scientific publications or sin- gle news article (Frank et al., 1999; Turney, 2000; Wan and Xiao, 2008; Jiang et al., 2009). We also notice that, currently, many websites provide the service which group related news together to fa- cilitate users’ browsing. In this paper, we focus on extracting keyphrases from a group of news arti- cles which describe the same news event by dif- ferent publishers. Previous studies on keyphrase extraction can be roughly categorized into two groups: super- vised and unsupervised. Unsupervised approaches usually select a set of candidates and use dif- ferent ranking methods to select the candidates with the highest scores as keyphrases. Most of ranking methods are based on the information ex- tracted from the document, such as TF·IDF, po- sition, syntactic relation with other words, and so on. Supervised methods convert the task into a binary classification problem, which categorizes phrases as keyphrases or non-keyphrases. Simi- lar as other tasks applied by supervised …

Content truncated to 500 words