Detecting Deceptive Utterances Using Deep Pre-Trained Neural Networks

Lying is an integral part of everyday communication in both written and oral forms. Detecting lies is therefore essential and has many possible applications. Our study aims to investigate the performance of automated lie detection methods, namely the most recent breed of pre-trained transformer neural networks capable of processing the Polish language. We used a dataset of nearly 1500 true and false statements, half of which were transcripts and the other half written statements, originating from possibly the largest study of deception in the Polish language. Technically, the problem was posed as text classification. We found that models perform better on typed than spoken utterances. The best-performing model achieved an accuracy of 0.69, which is much higher than the human performance average of 0.56. For transcribed utterances, human performance was at 0.58 and the models reached 0.62. We also explored model interpretability based on integrated gradient to shed light on classifier decisions. Our observations highlight the role of first words and phrases in model decisions, but more work is needed to systematically explore the observed patterns.