Abstract
We annotated emotion tags to text-dialogs in comics with focusing facial expressions in order to construct reliable dialog corpus with emotion tags, and evaluated the reliability of the constructed corpus. Generally, the relationship between language expression and emotion of the speaker is ambiguous, so it is difficult to distinguish correct emotion existing inside of the speaker with referring only the language expression in dialog. To solve this problem, there exist many investigations using non-lingual information of acoustic data though it costs to collect much speech data. Therefore, in this paper, we focuse on the facial expression appearing in comics to gain the reliability of the emotional annotating. For instance, we used 10 comic books on “Chibimarukochan” containing 29, 538 sentences and constructed emotional corpus annotated by the scheme, where two annotators annotate facial tags and emotion tags temporally and then decide correct tags by their discussion together. The correct emotion tags were 16, 635. The evaluation results proved that the agreement ratio of the temporary emotion tags between the two annotators was 78% which was as good as the related work in speech (83%) and the correctness of the decided emotion tags was 97%. Next, since in our trial experiment to extract emotional suffix expression from the corpus we successed to extract 3, 164 ones, the usability for dialog analysis on emotion was clarified. Thus, we confirmed that the facial expression in comics is as effective for distinction of emotions as speech and the schema of corpus construction using facial expression in comics is reliable.
