Prof. Zhang Yingqian from the School of Artificial Intelligence and Robotics at Xiamen University Malaysia, together with undergraduate student Zhou Chucheng (first author) and collaborators, recently published a research article in Scientific Reports, a peer-reviewed journal under the Nature Portfolio (SCI, JCR Q1, 2025 Impact Factor: 3.9).

The article, titled "A Deception Detection Model by Using Integrated LLM with Emotion Features", introduces a framework named LieXBerta, which aims to improve the accuracy and efficiency of lie detection in courtroom interrogations. Built on the RoBERTa large language model, LieXBerta integrates emotional signals such as anger, sadness, hopefulness, and hesitation with facial and action features. These multimodal inputs are processed through the Extreme Gradient Boosting (XGBoost) algorithm to provide objective and reliable deception analysis.

The study shows that LieXBerta achieved an accuracy of 87.5%, outperforming existing baseline models that rely only on facial or textual features. In addition, optimization reduced runtime by 42%, demonstrating the model’s efficiency and practical potential.

By offering more robust analysis of courtroom interrogation data, the research contributes to the development of intelligent judicial systems, with potential applications in supporting fair trials, reducing human bias, and strengthening transparency in legal processes. Their finding highlights the role of emotion-enhanced multimodal learning in advancing trustworthy AI for use in high-stakes decision-making contexts.

Image

https://doi.org/10.1038/s41598-025-17741-4