Evaluating text classification: A benchmark study

Manon Reusens, Alexander Stevens, Jonathan Tonglet, Johannes De Smedt, Wouter Verbeke, Seppe vanden Broucke, Bart Baesens

Abstract: Highlights•Overall, Bidirectional LSTMs (BiLSTMs) are the best-performing method.•LR with TF-IDF shows statistically similar results to BiLSTM and RoBERTa.•For fake news and topic detection, simple techniques are preferred.•The sentiment analysis tasks prefer more complex methods.•Smallest datasets prefer less complex techniques.