Applying Graph-based Keyword Extraction to Document Retrieval

International Joint Conference on Natural Language Processing, pages 864–868, Nagoya, Japan, 14-18 October 2013. Applying Graph-based Keyword Extraction to Document Retrieval Youngsam Kim1 Munhyong Kim1 Andrew Cattle1 Julia Otmakhova1 Suzi Park1 Hyopil Shin1 1Seoul National University/ Gwanak-1, Gwanak-ro, Gwanak-gu, Seoul, South Korea youngsamy@gmail.com, likerainsun@gmail.com, acattle@gmail.com, julia.nixie@gmail.com, mam3b@snu.ac.kr, hpshin@snu.ac.kr Abstract This paper proposes a keyword extraction process, based on the PageRank algorithm, to reduce noise of input data for measur- ing semantic similarity. This paper will in- troduce several features related to imple- mentation and discuss their effects. It will also discuss experimental results which showed significantly improved document retrieval performance with this extraction process in place. 1 Introduction To date, the most popular and well known ap- proach to calculating semantic similarity of text documents has been utilizing Vector Space Mod- els (VSM). The key idea of VSM is to map each document in a corpus or collection into a vector of a vector space, and interpret the distance be- tween the query document’s vector and the other texts’ vectors as their degree of semantic related- ness (Salton, 1971; Salton et al., 1975; Turney and Pantel, 2010). The VSM evolved from the SMART infor- mation retrieval system (Salton, 1971) and SMART pioneered many important terms and concepts that were adopted by modern search engines (Turney and Pantel, 2010). Many search engines are reported to use VSM to calculate the similarity between a query and a document (Manning et al., 2008). A common problem of VSM is that documents often contain words with high frequency but lit- tle semantic significance. VSM usually deal with this obstacle using tf-idf weighting and singular value decomposition techniques (Turney and Pantel, 2010). In order to improve retrieval systems that use the models, we suggest that employing keyword extraction on a per document basis to help reduce the noise inherent in large texts. For this purpose, PageRank, a graph-based ranking algorithm, was used in this study. Graph-based analysis tech- niques represent a document or text as a graph consisting of nodes (terms or phrases) and edges (pre-defined relations). Along with Brin and Page (1998)’s PageRank, various modifications and other graph-based algorithms have been in- troduced and proved their usefulness in various natural language processing tasks (Erkan and Radev, 2004; Kurland and Lee, 2006; Mihalcea et al., 2004; Wang et al., 2007; Widdows and Dorow, 2002). Graph-based extraction systems showed better performance over frequency-based systems on multiple-theme documents (Grineva et al., 2009). In this study, it was assumed that applications would benefit from being able to select important words from documents using the extraction sys- tem; not just for keyword extraction tasks, but also for any complex system that needs its input- data to be noise-reduced for future processes. 2 Theoretical Background In this paper, typical core parts of VSM are ap- plied to measure semantic similarity over docu- ments. Therefore, instead of using raw frequen- cies tf-idf weighting is adopted and length nor- malization is performed on both queries and tar- get documents (Salton and Buckley, …

Content truncated to 500 words