Human Values Estimation on News Articles through BERT-extracted Opinion Expressions

This paper proposes a human values estimation method with an opinion extraction approach. We assume that opinion sentences reflect human values the most. For a given article, the proposed method extracts the opinion sentences from the texts and estimates the human values included in the opinion sentences. The opinion sentence extraction is conducted by classifying each sentence as an opinion sentence or a non-opinion sentence. The proposed method concatenates sentences from the same article to extend the input texts as an upsampling approach while estimating the human values. The upsampling approach enriches the information volume of the training data. The distribution of human values from each opinion sentence shows the article's general human values. We conducted two evaluation experiments. We used the editorial articles from Mainichi Newspaper as the corpus data. The first experiment evaluated the performance of opinion sentence extraction. The training accuracy of opinion sentence extraction was 92%. In the evaluation test, the model reached a
$F_{1}$
score of 85.5%. The results showed that opinion sentences could be extracted with high accuracy. The second experiment evaluated the performance of human values estimation. There are five categories for human values estimation. The experiment was conducted with the same editorial articles from Mainichi Newspaper as the corpus data. We applied a training data enhancement approach by increasing the sentence number of training input and achieved a training accuracy up of over 50%. The results showed that the human values of opinion sentences could be estimated with high accuracy.