# Explainable Verbal Deception Detection using Transformers

**Authors:** Loukas Ilias, Felix Soldner, Bennett Kleinberg  
**arXiv:** [2210.03080](https://arxiv.org/abs/2210.03080)  
**Date:** 2022-10-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2210.03080v1/

## Key points

- Deceptive statements in daily life: fake news, misleading product reviews, lies about activities
- Limited use of deep learning approaches in automated text-based deception detection
- Lack of interpretability in deep learning methods
- Recent advancements allow for explaining aspects of deep learning models
- Proposal and evaluation of six deep-learning models utilizing BERT, MultiHead Attention, co-attentions, and transformers
- Use of LIME to examine model predictions and gain insights into decision-making process
- Focus on vocabulary uniqueness and correlation between LIWC categories and outcome class (truthful vs deceptive)
- Transformer-based models improve deception detection performance by 2.11% in accuracy
- Significant differences observed in LIWC feature usage between truthful and deceptive statements
- Enhances understanding of deep learning approaches for deception detection
- Transformer-based models show promise in effectively detecting deception in text-based content

## Summary

This paper titled "Explainable Verbal Deception Detection using Transformers" addresses the issue of deceptive statements that people encounter in their daily lives, such as fake news, misleading product reviews, or lies about activities. The authors highlight the limited use of deep learning approaches in automated text-based deception detection and the lack of interpretability in these methods. However, recent advancements have made it possible to explain certain aspects of deep learning models. The paper proposes and evaluates six deep-learning models that utilize BERT (and RoBERTa), MultiHead Attention, co-attentions, and transformers. To gain insights into how these models make decisions, the authors employ LIME to examine the model's predictions. Additionally, they focus on vocabulary uniqueness and the correlation between LIWC categories (linguistic inquiry and word count) and the outcome class (truthful vs deceptive). The findings indicate that transformer-based models can improve automated deception detection performance by 2.11% in accuracy. Moreover, significant differences are observed regarding the usage of LIWC features in truthful and deceptive statements. Overall, this research contributes to enhancing our understanding of deep learning approaches for deception detection while providing insights into how these models reach their decisions. The results suggest that transformer-based models show promise in effectively detecting deception in text-based content.