[4A1-GS-6] Language media processing:

[4A1-GS-6-03] Construction of a Video Translation Dataset with Added Character Personality and Interpersonal Relationship Information

〇Masatoshi Otake¹, Daichi Kato¹, Yuto Nozaki¹, Satoshi Hirooka¹, Yusuke Miyao¹, Asako Kanezaki² (1. The University of Tokyo, 2. Tokyo Institute of Technology)

Keywords:translation, NLP

This paper proposes a method for generating a dataset for video machine translation that includes information on the personality traits and relationships of characters appearing in visual media. It is known that in video translation, translators consider not only the textual information of the script but also meta information of the work, such as characters' personalities and relationships. However, such an approach has not been sufficiently explored in video machine translation. Therefore, this study proposes a method to clean data from external sources, including scripts and subtitles, and organize the information on speakers and their lines. This process involves separating and arranging the speaker's metadata and names, resulting in a Japanese-English parallel translation dataset that includes this detailed information. The data constructed by this method has been confirmed to have sufficient accuracy.