Representing Topics Using Images

Proceedings of NAACL-HLT 2013, pages 158–167, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Representing Topics Using Images Nikolaos Aletras and Mark Stevenson Department of Computer Science University of Sheffield Regent Court, 211 Portobello Sheffield, S1 4DP, UK {n.aletras, m.stevenson}@dcs.shef.ac.uk Abstract Topics generated automatically, e.g. using LDA, are now widely used in Computational Linguistics. Topics are normally represented as a set of keywords, often the n terms in a topic with the highest marginal probabilities. We introduce an alternative approach in which topics are represented using images. Candi- date images for each topic are retrieved from the web by querying a search engine using the top n terms. The most suitable image is se- lected from this set using a graph-based al- gorithm which makes use of textual informa- tion from the metadata associated with each image and features extracted from the images themselves. We show that the proposed ap- proach significantly outperforms several base- lines and can provide images that are useful to represent a topic. 1 Introduction Topic models are statistical methods for summaris- ing the content of a document collection using latent variables known as topics (Hofmann, 1999; Blei et al., 2003). Within a model, each topic is a multino- mial distribution over words in the collection while documents are represented as distributions over top- ics. Topic modelling is now widely used in Natural Language Processing (NLP) and has been applied to a range of tasks including word sense disambigua- tion (Boyd-Graber et al., 2007), multi-document summarisation (Haghighi and Vanderwende, 2009), information retrieval (Wei and Croft, 2006), image labelling (Feng and Lapata, 2010a) and visualisation of document collections (Chaney and Blei, 2012). Topics are often represented by using the n terms with the highest marginal probabilities in the topic to generate a set of keywords. For example, wine, bot- tle, grape, flavour, dry. Interpreting such lists may not be straightforward, particularly since there may be no access to the source collection used to train the model. Therefore, researchers have recently begun developing automatic methods to generate meaning- ful and representative labels for topics. These tech- niques have focussed on the creation of textual la- bels (Mei et al., 2007; Lau et al., 2010; Lau et al., 2011). An alternative approach is to represent a topic us- ing an illustrative image (or set of images). Im- ages have the advantage that they can be under- stood quickly and are language independent. This is particularly important for applications in which the topics are used to provide an overview of a collec- tion with many topics being shown simultaneously (Chaney and Blei, 2012; Gretarsson et al., 2012; Hinneburg et al., 2012). This paper explores the problem of selecting im- ages to illustrate automatically generated topics. Our approach generates a set of candidate images for each topic by querying an image search engine with the top n topic terms. The most suitable image is selected using a graph-based method that makes use of both textual and visual information. Textual …

Content truncated to 500 words