Bias Detection and Mitigation in Sentiment Analysis

Natural language processing applications play an important role in our daily life. Even if the success of these applications is high, the decisions they take may not be fair and may contain bias. Especially in decision making systems that affect human life, it is important that this application be fair (not biased towards individuals or groups) and explainable. There may be many resources of bias such as the training data set, the algorithm which is used or the trained model. In this study, gender bias in the data set in sentiment analysis was examined. In the study, both Turkish and English data sets were used and the results obtained for the two languages were compared. And also in the study, the results of different feature extraction methods in the test results were compared by using different feature extraction methods with classical machine learning. In addition to these, in the study, the reason for this bias was investigated and it was studied to mitigate it. Comparative results were obtained by using classical machine learning together with word2vec and fasttext for bias detection. Studies have been made to reduce bias using Google's text-to-text transfer model T5 and LSTM (long short-term memory) models.