Is Similarity Visually Grounded? Computational Model of Similarity for the Estonian language

Proceedings of Recent Advances in Natural Language Processing, pages 541–549, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_064 541 Is Similarity Visually Grounded? Computational Model of Similarity for the Estonian language Claudia Kittask Institute of Computer Science University of Tartu Tartu, Estonia claudiakittask@gmail.com Eduard Barbu Institute of Computer Science University of Tartu Tartu, Estonia eduard.barbu@ut.ee Abstract Researchers in Computational Linguistics build models of similarity and test them against human judgments. Although there are many empirical studies of the compu- tational models of similarity for the En- glish language, the similarity for other lan- guages is less explored. In this study we are chiefly interested in two aspects. In the first place we want to know how much of the human similarity is grounded in the visual perception. To answer this question two neural computer vision mod- els are used and their correlation with the human derived similarity scores is com- puted. In the second place we investigate if language influences the similarity com- putation. To this purpose diverse com- putational models trained on Estonian re- sources are evaluated against human judg- ments. 1 Introduction Various disciplines and research communities are interested in the study of similarity: Philosophy, Psychology, Computational Linguistics, Semantic Web and the Linked Data communities, to name a few. For example, to integrate heterogeneous semantic resources the Linked Data and Seman- tic Web communities estimate the degree of sim- ilarities between the concepts in these resources (Euzenat and Shvaiko, 2013; Harispe et al., 2015). In Computational Linguistics researchers build computational models of similarity. To test them, the correlation between the human similar- ity scores and the scores assigned by the computa- tional models is calculated. It is assumed that the best computational models predict better the hu- man judge scores. However, earlier studies of similarity suffered from a drawback: they do not distinguish be- tween the relations of similarity and association. In psychology, for example, the distinction be- tween these two notions is well understood. The association between two concepts is defined as the propensity of a subject to activate a representa- tion of the second concept when the first concept is presented. In contrast, the similarity is defined as the proximity of two mental representations. In the Gestalt psychology for example (Wertheimer, 1938) the similarity is seen as the principle of or- ganization of objects in perceptual groups. The concepts cup and tea are associated but not sim- ilar: there is no perceptual principle to group to- gether an object like a cup and a liquid like tea. However, the objects denoted by the concepts ap- ple and pear are perceptually similar. To remedy this problem SimLex-999 (Hill et al., 2015), a gen- uine data similarity set containing human judge similarity and concreteness scores for 999 English word pairs, has been built. An interesting distinction is that between the surface similarity and deep similarity (Vosniadou and Ortony, 1989). The surface similarity is per- ceptually grounded and it is used in categorization. In contrast, the deep similarity is related to …

Content truncated to 500 words