Improving Emotion Classification by a Combination of Personal Texts and Social Big Data Based on Naive Bayes

Nowadays, a lot of personal texts are available due to the rise of social networking platforms. Using the texts, many research are conducted to estimate emotions. Especially, the tweets posted on Twitter are used in research employing machine learning as a training data for emotion analysis and classification. In the classification, it may be possible to improve the accuracy by using one’s own past texts as a training data. In this study, we combine personal texts and Twitter data as the training data and try to improve the accuracy of emotion estimation by using Naive Bayes. We obtain texts from one’s past diary. In addition, tweets are collected from similar attributes accounts. By an evaluator, the texts and tweets are classified into three categories: positive, negative, and neutral. Then, we set four patterns by changing the combination of personal texts and tweets. As a result, the highest average accuracy was obtained by weighting words that appeared in personal texts and combining them with Twitter data. It shows that it is important to focus on words that the person has used in the past in estimating his/her own emotions.