Cross-Lingual Analysis of Semantic Networks Built on Word2Vec and Google N-Grams

In distributional semantics a huge variety of methods and tools are developed for categorizing semantic similarities. Nevertheless, the application of network analysis to the study of synonymy can expand. The goal of this work is to create semantic networks (graphs) built on both Word2Vec model and data from Google N-Grams, apply some metrics which were elaborated by the graph theory and compare the results for the graphs of different languages. The graphs for Word2Vec and Google N-Grams will be built differently. For Word2Vec models cosine similarity will be calculated, whereas for Google N-grams – bigram frequency. Consequently, the graphs will represent semantic distribution of the words but in different ways: Word2Vec networks will link synonyms, antonyms, hypernyms, hyponyms, Google N-Grams networks – the ‘neighbours’ of the words. As a result, three Google 2-Grams networks (for German, Italian, and Russian) will be created and two word2vec networks (for Russian and Italian) and the applied graph metrics will be linguistically interpreted.