Enriching Dictionaries with Images from the Internet - Targeting Wikipedia and a Japanese Semantic Lexicon: Lexeed -

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 331–339, Beijing, August 2010 Enriching Dictionaries with Images from the Internet - Targeting Wikipedia and a Japanese Semantic Lexicon: Lexeed - Sanae Fujita NTT Communication Science Lab. sanae@cslab.kecl.ntt.co.jp Masaaki Nagata NTT Communication Science Lab. nagata.masaaki@lab.ntt.co.jp Abstract We propose a simple but effective method for enriching dictionary definitions with images based on image searches. Vari- ous query expansion methods using syn- onyms/hypernyms (or related words) are evaluated. We demonstrate that our method is effective in obtaining high- precision images that complement dictio- nary entries, even for words with abstract or multiple meanings. 1 Introduction The Internet is an immense resource for images. If we can form connections between these im- ages and dictionary definitions, we can create rich dictionary resources with multimedia infor- mation. Such dictionaries have the potential to provide educational (Popescu et al., 2006), cross- langauge information retrieval (Hayashi et al., 2009) or assistive communication tools especially for children, language learners, speakers of differ- ent languages, and people with disabilities such as dyslexia (Mihalcea and Leong, 2008; Goldberg et al., 2009). Additionally, a database of typical images con- nected to meanings has the potential to fill the gaps between images and meanings (semantic gap). There are many studies which aim to cross the semantic gap (Ide and Yanai, 2009; Smeulders et al., 2000; Barnard et al., 2003) from the point of view of image recognition. However the se- mantic classes of target images are limited (e.g. Caltech-101, 2561). Yansong and Lapata (2008) tried to construct image databases annotated with keywords from Web news images with their cap- tions and articles, though the semantic coverage is 1http://www.vision.caltech.edu/Image Datasets/Caltech101, 256/ unknown. In this paper, we aim to supply sev- eral suitable images for dictionary definitions. We propose a simple but effective method based on an Internet image search. There have been several studies related to sup- plying images for a dictionary or thesaurus. Bond et al. (2009) applied images obtained from the Open Clip Art Library (OCAL) to Japanese Word- Net.2 They obtained candidate images by compar- ing the hierarchical structures of OCAL and Word- Net, and then judged whether or not the image was suitable for the synset by hand. OCAL benefits from being in the public domain; however, it can- not cover a wide variety of meanings because of the limited number of available images. Fujii and Ishikawa (2005) collected images and text from the Internet by querying lemma, and linked them to an open encyclopedia, CY- CLONE.3 They guessed the meaning of the im- ages by disambiguating the surrounding text. This is a straightforward approach, but it is difficult to use it to collect images with minor meanings, be- cause in most cases the Internet search querying lemma only provides images related to the most common meaning. For example, lemma y�� arch may mean ‘‘architecture’’ or ‘‘home run’’ in Japanese, but a lemma search provided no im- age of the latter at least in the top …

Content truncated to 500 words