A deception detection model by using integrated LLM with emotion features

Traditional lie detection relies on the experience of human interrogators, making it susceptible to subjective factors and leading to misjudgments. To solve this problem, we propose an emotion-enhanced deception detection model, Lie Detection using XGBoost with RoBERTa-based Emotion Features (LieXBerta). In this framework, the Robustly Optimized BERT Pretraining Approach (RoBERTa) is used to extract emotional features from interrogation texts. The emotional features are then combined with facial and action features and subsequently fed into an Extreme Gradient Boosting (XGBoost) classifier for deception detection. This approach aims to improve the objectivity and accuracy of deception detection in courtroom settings. For verifying the proposed algorithm, we develop a trial text dataset enriched with detailed emotional features. Simulation experiments demonstrate that the LieXBerta model incorporating emotional features outperforms baseline models that use only traditional features and several classical machine learning models. The experimental results show that after parameter tuning, the accuracy of the LieXBerta model increased to 87.50%, respectively, marking a 6.5% improvement over the baseline model. Moreover, the runtime of the tuned LieXBerta model with reduced features was reduced by 42%, significantly enhancing the training efficiency and prediction performance for deception detection.