Cross-Lingual Generation and Evaluation of a Wide-Coverage Lexical Semantic Resource

Cross-Lingual Generation and Evaluation of a Wide-Coverage Lexical Semantic Resource Attila Novák, Borbála Novák Pázmány Péter Catholic University Faculty of Information Technology and Bionics MTA-PPKE Hungarian Language Technology Research Group 1083 Budapest, Práter u. 50/a, Hungary {novak.attila, novak.borbala}@itk.ppke.hu Abstract Neural word embedding models trained on sizable corpora have proved to be a very efficient means of representing meaning. However, the abstract vectors representing words and phrases in these models are not interpretable for humans by themselves. In this paper we present the Thing Recognizer, a method that assigns explicit symbolic semantic features from a finite list of terms to words present in an embedding model, making the model interpretable for humans and covering the semantic space by a controlled vocabulary of semantic features. We do this in a cross-lingual manner, applying semantic tags taken form lexical resources in one language (English) to the embedding space of another (Hungarian). Keywords: semantic lexicon induction, word embedding models, cross-lingual resource generation 1. Introduction A recently very popular and efficient method for the dis- tributional representation of words is using word embed- ding (WE) models (Mikolov et al., 2013c). In this paper we present a method that creates the WE model of a large text corpus and inserts the corresponding embedding vectors of a limited set of abstract semantic features into the same space. The embedding vectors for semantic features are built from automatically reorganized lexical resources (that may be in a language different from our target language) and are transformed to the target WE space. Then, a near- est neighbor approach is applied to find the most relevant features for a query word. The assigned features can also be used as a searchable semantic annotation of the original corpus the WE model was created from, because our model assigns semantic features to any (even non-standard/slang or misspelled) word in a text in a language-independent manner, regardless of whether these are present in a lexical resource or not, and whether any such resource is available for the target language. The organization of categories and the way they are actually assigned to words by the algo- rithm is in accordance with the actual usage of these words as manifested by their distribution in a large corpus. The method is demonstrated for English and Hungarian, but it can easily be applied to other languages as well. 2. Related Work WE models have frequently been used to represent word meaning efficiently (Mikolov et al., 2013b; Pennington et al., 2014). There are also approaches that replace WE with sense embedding(Bordes et al., 2012; Neelakantan et al., 2014; Tian et al., 2014; Li and Jurafsky, 2015; Bartunov et al., 2015). Huang et al. (2012) applied clustering algo- rithms to create single prototype embedding. Some have tried to match WE’s to entities in existing lexi- cal resources, for example to BabelNet entries (Panchenko, 2016) or WordNet synsets (Chen et al., 2014; Agirre et al., 2006). Rothe and Schütze (2015) combines WE vec- tors to obtain Wordnet synset representations in the …

Content truncated to 500 words