Ideology Detection Using Transformer-Based Machine Learning Models

Ideology detection has been a challenging but essential problem that has been studied for a long time. Certain groups and organizations, such as politicians, rely on people’s political views to make wise or forward-looking decisions. In the previous days, intensive survey studies were needed to collect the opinions of the people, and it was a very laborious and challenging process to analyze the political tendencies of the citizens. For example, it was observed that the answers given to the questions asked when people were not anonymous while participating in the surveys were biased or abstained. For this reason, the results of classical survey methods are open to dis- cussion. Today, many users on social media have become accessible data sources and are used extensively in political research. On social media platforms such as Twitter, people reflect their political views with the comments they share. Inspired by recent studies that have successfully modeled sentiment analysis, we use a dataset called the Ideological Books Corpus (IBC) to identify emerging political opinion in a sentence. By using natural language processing methods, we delete unnecessary words, punctuation marks and apply Bidirectional Encoder Representations from Transformers (BERT), Long Short Term Memory (LSTM), Support Vector Machine, Decision Trees (Decision Trees), and Naive Bayes Classifier methods. As a result, it has been ob- served that the ELECTRA outperforms the other approaches we have used in terms of the F1-score performance metric.