Abstract
Distributed word representations in natural language processing have been used to estimate cognitive information from sentences (e.g., sentiment analysis). However, such estimation is difficult for highly subjective contents. We here propose a new technique that improves the performance in such estimation by incorporating human brain information into distributed word representations. In this technique, distributed representations, of given sentences, transformed into brain-activity representations are used to estimate the cognitive information linked with the sentences. To verify our technique, we obtained distributed word representations from a text corpus via word2vec, and modeled the transformation of the distributed representations into brain-activity representations using movie-evoked brain activity measured with functional MRI. We then applied our technique to the estimation of impression and preference for movie scenes from manual descriptions for the same scenes. We found that the performance of these estimations was higher when we used brain-activity representations transformed from distributed representations than when we used the distributed representations directly. This result suggests that distributed word representations can be improved by incorporating human brain information into them.
