Composing Word Vectors for Japanese Compound Words Using Dependency Relations

Kanako Komiya, Takumi Seitou, Minoru Sasaki, Hiroyuki Shinnou

Abstract: The use of distributed representations, e.g., via word2vec, has become popular in recent years. However, Japanese has many compound words and we often face the situation where meanings of a word and a compound word should be compared. Therefore, in the current study, we composed compound word vectors from those of constituent word vectors. We took into consideration the dependency relations of compound words to compose word vectors of them. The experiments revealed that, when we consider dependency relations, (1) we could obtain better representations for compound words when we separately learn models for each dependency relation, (2) each model could obtain good representations with fewer epochs, and (3) the learned weights for a model of compound words with one dependency relation could be used for fine-tuning for models for compound words of other dependency relations.