A Proposal of Emotion Estimation Method for Words and Construction of Word-emotion Dictionary

Abstract

: In this paper, we propose an emotion estimation method for words and its application to construct a word-emotion dictionary. Since the estimation of more delicate emotions in various natural language processing tasks is a crucial issue, we aim to estimate various emotions to a word. In order to realize it, we employ the distributional hypothesis and assume that an emotional word in a sentence influences the surrounding words. First, we collect more than 2,000 words expressing emotions from an emotion expression dictionary. By using these emotional words and a neural model based on Continuous Bag-of- Words (CBOW), we propose an automatic system to estimate the emotions of many ordinary words. As a result, emotion vectors for 20,000 words could be obtained. We carried out experiments to examine the accuracy of the vectors. It is confirmed that the generated emotion vectors reflect the emotion image for words that humans have.