Experimental Comparison of Pre-Trained Word Embedding Vectors of Word2Vec, Glove, FastText for Word Level Semantic Text Similarity Measurement in Turkish

Abstract

This study aims to evaluate experimentally the word vectors produced by three widely used embedding methods for the word-level semantic text similarity in Turkish. Three benchmark datasets SimTurk, AnlamVer, and RG65_Turkce are used in this study to evaluate the word embedding vectors produced by three different methods namely Word2Vec, Glove, and FastText. As a result of the comparative analysis, Turkish word vectors produced with Glove and FastText gained better correlation in the word level semantic similarity. It is also found that The Turkish word coverage of FastText is ahead of the other two methods because the limited number of Out of Vocabulary (OOV) words have been observed in the experiments conducted for FastText. Another observation is that FastText and Glove vectors showed great success in terms of Spearman correlation value in the SimTurk and AnlamVer datasets both of which are purely prepared and evaluated by local Turkish individuals. This is another indicator showing that these aforementioned datasets are better representing the Turkish language in terms of morphology and inflections.