Automatic Keyphrase Extraction via Topic Decomposition

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 366–376, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Automatic Keyphrase Extraction via Topic Decomposition Zhiyuan Liu, Wenyi Huang, Yabin Zheng and Maosong Sun Department of Computer Science and Technology State Key Lab on Intelligent Technology and Systems National Lab for Information Science and Technology Tsinghua University, Beijing 100084, China {lzy.thu, harrywy, yabin.zheng}@gmail.com, sms@tsinghua.edu.cn Abstract Existing graph-based ranking methods for keyphrase extraction compute a single impor- tance score for each word via a single ran- dom walk. Motivated by the fact that both documents and words can be represented by a mixture of semantic topics, we propose to decompose traditional random walk into mul- tiple random walks specific to various topics. We thus build a Topical PageRank (TPR) on word graph to measure word importance with respect to different topics. After that, given the topic distribution of the document, we fur- ther calculate the ranking scores of words and extract the top ranked ones as keyphrases. Ex- perimental results show that TPR outperforms state-of-the-art keyphrase extraction methods on two datasets under various evaluation met- rics. 1 Introduction Keyphrases are defined as a set of terms in a doc- ument that give a brief summary of its content for readers. Automatic keyphrase extraction is widely used in information retrieval and digital library (Tur- ney, 2000; Nguyen and Kan, 2007). Keyphrase ex- traction is also an essential step in various tasks of natural language processing such as document cate- gorization, clustering and summarization (Manning and Schutze, 2000). There are two principled approaches to extracting keyphrases: supervised and unsupervised. The su- pervised approach (Turney, 1999) regards keyphrase extraction as a classification task, in which a model is trained to determine whether a candidate phrase is a keyphrase. Supervised methods require a doc- ument set with human-assigned keyphrases as train- ing set. In Web era, articles increase exponentially and change dynamically, which demands keyphrase extraction to be efficient and adaptable. However, since human labeling is time consuming, it is im- practical to label training set from time to time. We thus focus on the unsupervised approach in this study. In the unsupervised approach, graph-based rank- ing methods are state-of-the-art (Mihalcea and Ta- rau, 2004). These methods first build a word graph according to word co-occurrences within the docu- ment, and then use random walk techniques (e.g., PageRank) to measure word importance. After that, top ranked words are selected as keyphrases. Existing graph-based methods maintain a single importance score for each word. However, a docu- ment (e.g., news article or research article) is usu- ally composed of multiple semantic topics. Taking this paper for example, it refers to two major top- ics, “keyphrase extraction” and “random walk”. As words are used to express various meanings corre- sponding to different semantic topics, a word will play different importance roles in different topics of the document. For example, the words “phrase” and “extraction” will be ranked to be more impor- tant …

Content truncated to 500 words