Automatic generation of inter-passage links based on semantic similarity

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 590–598, Beijing, August 2010 Automatic generation of inter-passage links based on semantic similarity Petr Knoth, Jakub Novotny, Zdenek Zdrahal Knowledge Media Institute The Open University p.knoth@open.ac.uk Abstract This paper investigates the use and the prediction potential of semantic similar- ity measures for automatic generation of links across different documents and pas- sages. First, the correlation between the way people link content and the results produced by standard semantic similarity measures is investigated. The relation be- tween semantic similarity and the length of the documents is then also analysed. Based on these findings a new method for link generation is formulated and tested. 1 Introduction Text retrieval methods are typically designed to find documents relevant to a query based on some criterion, such as BM25 or cosine similar- ity (Manning et al., 2008). Similar criteria have also been used to identify documents relevant to the given reference document, thus in principle linking the reference document to the related doc- uments (Wilkinson and Smeaton, 1999). This pa- per studies the correspondence between the results of this approach and the way linking is performed by people. The study confirms that the length of documents is an important factor usually causing the quality of current link generation approaches to deteriorate. As a result, methods working at a finer granularity than documents should be in- vestigated. This will also improve the speed of access to information. For example, when users read through a long document, they should be able to quickly access a passage in another possibly This work has been partially supported by Eurogene - Contract no. ECP-2006-EDU-410018) long document related to the discussed topic. The automatic detection of document pairs containing highly related passages is the task addressed in this paper. A number of approaches for automatic link generation have used measures of semantic sim- ilarity. While these measures were widely used for the discovery of related documents in prac- tise, their correspondence to the way people link content has not been sufficiently investigated (see Section 2). As our contribution to this topic, we present in this paper an approach which tries to first investigate this correspondence on a large text corpus. The resulting method is then motivated by the outcomes of this analysis. It has been recognised in information retrieval that when a collection contains long documents, better performance is often achieved by breaking each document into subparts or passages and com- paring these rather than the whole documents to a query (Manning et al., 2008). A suitable granular- ity of the breakdown is dependent on a number of circumstances, such as the type of the document collection or the information need. In this work, we have decided to work at the level of documents and paragraphs. Our task can be formalized as a two-step process: 1. Given a collection of documents, our goal is to identify candidate pairs of documents be- tween which a link may be induced. 2. …

Content truncated to 500 words