[4Rin1] Interactive 2

[4Rin1-01] Relationship between proportional relationship in stationary probabilities and analogical reasoning

〇Akimichi Kamei¹, Shohei Hidaka¹, Torii Takuma¹ (1.Japan Advanced Institute of Science and Technology)

Keywords:Semantic space, Analogy, Markov process

A word embedding model such as Word2vec is supposed to embed a high-dimensional word-to-word co-occurrence probabilities into a low-dimensional vector space. Past theoretical studies on word2vec have concluded this embedding is important to form a vector space, which can be useful for semantic tasks such as analogical reasoning. In this study, however, we revisited this conclusion, and re-analyzed statistical nature of the co-occurrence matrix, rather than word2vec. As a result of this theoretical exploration, we found that the relationship between a kind of proportional relationship in transition probability and that in stationary probability, and also found this theoretical relationship was related to the performance in analogical reasoning. This finding leads us to consider the direct relationship between underlying statistical nature of word-to-word co-occurrence and analogical reasoning.