Analogy comprehension between psychological experiments and word embedding models

抄録

Word embedding models such as word2vec (Mikolov,2013) and GloVe (Pennington et al.2014) became to be widely recognized, while similarity judgment of analogy and comprehension could be described on a Euclidian space (Rumelhart and Abrahamson, 1973). We tried to compare between word embedding models and human performances. Despite that several models were proposed to deal with adjusting each domain in order to improve their performance, it has still remained unsolved the relation between human judgment and understanding of analogy and their performance of word embedding models. We investigated the relationship between them. Our results suggest that human understanding of analogy might be understood in terms of word embedding spaces. We discussed the further possibilities to understand the semantic space in human in these models.