Towards Semantic Language Classification: Inducing and Clustering Semantic Association Networks from Europarl

Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics (*SEM 2015), pages 127–136, Denver, Colorado, June 4–5, 2015. Towards Semantic Language Classification: Inducing and Clustering Semantic Association Networks from Europarl Steffen Eger1, Niko Schenk2 and Alexander Mehler1 1Text Technology Lab 2Applied Computational Linguistics Lab Goethe University Frankfurt am Main {steeger,nschenk,amehler}@em.uni-frankfurt.de Abstract We induce semantic association networks from translation relations in parallel corpora. The resulting semantic spaces are encoded in a single reference language, which ensures cross-language comparability. As our main contribution, we cluster the obtained (cross- lingually comparable) lexical semantic spaces. We find that, in our sample of languages, lexical semantic spaces largely coincide with genealogical relations. To our knowledge, this constitutes the first large-scale quantita- tive lexical semantic typology that is com- pletely unsupervised, bottom-up, and data- driven. Our results may be important for the decision which multilingual resources to inte- grate in a semantic evaluation task. 1 Introduction There has been a recent surge of interest in integrat- ing multilingual resources in natural language pro- cessing (NLP). For example, Snyder et al. (2008) show that jointly considering morphological seg- mentations across languages improves performance compared to the monolingual baseline. Bhargava and Kondrak (2011) and Bhargava and Kondrak (2012) demonstrate that string transduction can ben- efit from supplemental information provided in other languages. Analogously, in lexical semantics, Nav- igli and Ponzetto (2012) explore semantic relations from Wikipedia in different languages to induce a huge integrated lexical semantic network. In this paper, we also focus on multilingual re- sources in lexical semantics. But rather than inte- grating them, we investigate their (dis-)similarities. More precisely, we cluster (classify) languages based on their semantic relations between lexical units. The outcome of our classification may have direct consequences for approaches that integrate di- verse multilingual resources. For example, from a linguistic point of view, it might be argued that in- tegrating very heterogeneous/dissimilar semantic re- sources is harmful, e.g., in a monolingual semantic similarity task, because semantically unrelated lan- guages might contribute semantic relations unavail- able in the language for which semantic similarity is computed. Alternatively, from a statistical point of view, it might be argued that integrating heteroge- neous/dissimilar resources is beneficial due to their higher degree of uncorrelatedness. In any case, ei- ther of these implications necessitates knowledge of a typology of lexical semantics. In order to address this question, we provide a translation-based model of lexical semantic spaces. Our approach is to generate association networks in which the weight of a link between two words de- pends on their degree of partial synonymy. To mea- sure synonymy, we rely on translation data that is input to a statistical alignment toolkit. We define the degree of synonymy of two words to be proportional to the number of common translations in a reference language, weighted by the probability of translation. By pivoting on the reference language, we represent semantic associations among words in different lan- guages by means of the synonymy relations of their translations in the same target language. …

Content truncated to 500 words