Identifying Important Features for Graph Retrieval

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 600–609, Dublin, Ireland, August 23-29 2014. Identifying Important Features for Graph Retrieval Zhuo Li and Sandra Carberry and Hui Fang* and Kathleen F. McCoy ivanka@udel.edu carberry@udel.edu hui@udel.edu mccoy@udel.edu Department Computer and Information Science, *Department of Electrical and Computer Engineering University of Delaware Abstract Infographics, such as bar charts and line graphs, occur often in popular media and are a rich knowledge source that should be accessible to users. Unfortunately, information retrieval re- search has focused on the retrieval of text documents and images, with almost no attention specif- ically directed toward the retrieval of information graphics. Our work is the first to directly tackle the retrieval of infographics and to design a system that takes into account their unique charac- teristics. Learning-to-rank algorithms are applied on a large set of features to develop several models for infographics retrieval. Evaluation of the models shows that features pertaining to the structure and the content of graphics should be taken into account when retrieving graphics and that doing so results in a model with better performance than a baseline model that relies on matching query words with words in the graphic. 1 Introduction Infographics are non-pictorial graphics such as bar charts and line graphs. When such graphics appear in popular media, they generally have a high-level message that they are intended to convey. For example, the graphic in Figure 1 ostensibly conveys the message that Toyota has the highest profit among the automobile companies listed. Thus infographics are a form of language since, according to Clark (Clark and Curran, 2007), language is any deliberate signal that is intended to convey a message. Although much research has addressed the retrieval of documents, very little attention has been given to the retrieval of infographics. Yet research has shown that the content of an infographic is often not included in the article’s text (Carberry et al., 2006). Thus infographics are an important knowledge source that should be accessible to users of a digital library. Techniques that have been effective for document or image retrieval are inadequate for the retrieval of infographics. Current search engines employ strategies similar to those used in document retrieval, relying primarily on the text surrounding a graphic and web link structures. But the text in the surround- ing document generally does not refer explicitly to the infographic or even describe its content (Carberry et al., 2006). An obvious extension to using the article text would be to collect all the words in an infographic and use it as a bag of words. However, infographics have structure and often a high-level message, and bag of words approaches ignore this structure and message content. This paper explores the features that should be taken into account when ranking graphics for retrieval in response to a user query. Using a learning-to-rank algorithm on a wide range of features (including structural and content features), we produce a model that performs significantly better than …

Content truncated to 500 words