A deception detection model by using integrated LLM with emotion features

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Traditional lie detection methods heavily rely on human interrogators, introducing subjectivity and potential misjudgments.
• Developing objective and accurate deception detection systems is crucial for legal and forensic applications.

Purpose of the Study:

• To propose and evaluate an emotion-enhanced deception detection model, Lie Detection using XGBoost with RoBERTa-based Emotion Features (LieXBerta), for improved objectivity and accuracy.
• To integrate emotional, facial, and action features for more robust deception detection.

Main Methods:

• Utilized the Robustly Optimized BERT Pretraining Approach (RoBERTa) to extract emotional features from interrogation texts.
• Combined extracted emotional features with facial and action features.
• Employed an Extreme Gradient Boosting (XGBoost) classifier for the deception detection task.
• Developed a trial text dataset enriched with detailed emotional features for model verification.

Main Results:

• The LieXBerta model, incorporating emotional features, demonstrated superior performance compared to baseline models using only traditional features and other classical machine learning models.
• Parameter tuning resulted in an accuracy of 87.50% for the LieXBerta model, a 6.5% improvement over the baseline.
• Reduced features in the tuned LieXBerta model decreased runtime by 42%, enhancing training efficiency and prediction performance.

Conclusions:

• The proposed LieXBerta model significantly enhances the objectivity and accuracy of deception detection by integrating emotional features.
• The model shows promising potential for application in courtroom settings, offering a more reliable alternative to traditional methods.
• The integration of RoBERTa for emotion extraction and XGBoost for classification provides an effective framework for advanced deception detection.