Developing a Dataset for Mining Reviews in Tweets focusing on Celebrities' Aspects

Kotoe Sugawara, Takehito Utsuro

Abstract: The purpose of this paper is to make it easier for celebrity fans to search for information on critiques and interest trends about celebrities and matters related to them. As part of this process, we collect tweets on Twitter that mention aspects of celebrities and train and evaluate a token classification framework to judge whether or not tweets represent reviews on aspects of celebrities. Specifically, we develop a dataset by annotating the collected tweets whether or not aspect’s names of celebrities and adjectives/adjectival verbs have review relationships or not. We train and evaluate the token classification framework using BERT to determine whether the tweets represent reviews on aspects of celebrities. In the evaluation, we evaluate whether or not the TARGET and REVIEW labels annotated by the token classification framework using BERT are correct or not. The token classification framework using BERT achieved high enough performance, indicating that it is appropriate to apply the token classification framework using BERT to this task.