LegalEye: Multimodal Court Deception Detection Across Multiple Languages.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Linguistics

Background:

• Deception detection in legal settings is challenging.
• Multimodal data analysis offers potential for improved accuracy.
• Reducing demographic bias in AI models is crucial for fairness.

Purpose of the Study:

• Introduce LegalEye, a multimodal AI model for deception detection.
• Evaluate the impact of integrating audio, visual, and textual data.
• Assess bias mitigation strategies in legal AI.

Main Methods:

• Developed LegalEye using neural networks and late fusion.
• Analyzed courtroom testimony data in English, Spanish, and Tagalog.
• Constructed a balanced dataset across racial groups and genders.

Main Results:

• Achieved high deception detection rates: 97% (English), 85% (Spanish), 86% (Tagalog).
• Late fusion of modalities outperformed single-modality analysis.
• Visual features were most influential for English and Tagalog; audio/text for Spanish.

Conclusions:

• LegalEye demonstrates potential for language-adaptive, culturally sensitive deception detection.
• The model can support legal counsel and promote fairer judicial outcomes.
• Further research should expand linguistic and demographic diversity.