The Lie Detector: Explorations in the Automatic Recognition of Deceptive Language

Rada Mihalcea; Carlo Strapparava

In brief

Mihalcea and Strapparava investigate automatic classification techniques for distinguishing between truthful and deceptive written language. Using three datasets of labeled true and false statements on abortion, death penalty, and best friends, they achieve ~70% classification accuracy with Naive Bayes and SVM classifiers, significantly outperforming the 50% baseline. Analysis of linguistic features reveals that deceptive texts exhibit detachment from self (higher use of second/third person pronouns) while truthful texts show stronger self-identification and belief-oriented language.

Contributions

Introduces three manually-created datasets (100 true and 100 false statements each) on three topics, collected via Amazon Mechanical Turk with explicit ground truth labels for truth/deception
Demonstrates that truthful and deceptive language are separable using standard NLP classification techniques
Shows that the separability property generalizes across different topics through cross-topic transfer experiments
Develops a class-based feature dominance metric using LIWC word categories to identify salient linguistic features characteristic of deceptive texts

Method

The authors construct three balanced datasets (abortion opinions, death penalty opinions, best friend descriptions) where contributors write both truthful statements and deliberately deceptive statements in response to specific prompts (10-15 minutes to write 4-5+ detailed sentences). After collecting 100 true and 100 false statements per topic, they apply standard NLP preprocessing (tokenization, stemming, no feature selection, no stopword removal) and classify using two diverse classifiers: Naive Bayes and Support Vector Machines with 10-fold cross-validation.

To identify distinctive linguistic features, they define a dominance score metric: for each word class C, dominance is the ratio of class coverage in the deceptive corpus to class coverage in the truthful corpus. They apply this to LIWC (Linguistic Inquiry and Word Count), a validated psycholinguistic lexicon with ~70 categories of ~2,200 words, ranking word classes by dominance score to identify which linguistic patterns are characteristic of lies versus truth.

Results

Classification accuracy averaged 70% across both classifiers and all three topics (Naive Bayes 70.8%, SVM 70.1%), substantially above the 50% baseline. Performance varied by topic: best friend (75-77%), abortion (67.5-70%), death penalty (65.9-67.4%). Learning curves showed monotonic improvement with more training data, suggesting potential for better performance with larger datasets. Cross-topic transfer experiments (training on two topics, testing on the third) yielded 59.8% (Naive Bayes) and 57.8% (SVM) accuracy—below in-topic performance but still significantly above baseline, indicating the classifiers learned topic-general deception cues rather than topic-specific artifacts.

Analysis of LIWC word classes revealed striking patterns:

Deceptive text is dominated by:

Human-related categories (YOU 1.53, OTHER 1.47, HUMANS 1.31) indicating detachment from self
Certainty-related words (CERTAIN 1.24: "always", "completely", "truly") suggesting compensatory emphasis on truth
Metaphor (METAPH 1.71: "sacred", "soul", "mercy") suggesting abstraction

Truthful text is dominated by:

Self-related categories (I 0.59, SELF 0.64, FRIENDS 0.63) indicating comfortable self-identification
Belief-oriented language (INSIGHT 0.65: "believe", "think", "feel") reflecting genuine conviction
Optimism (OPTIM 0.57: "best", "hope")

Notes

This is an early computational approach to deception detection focused on explicit linguistic features accessible from short-form text. The strength is careful dataset construction with true ground truth (genuine opinions vs. deliberate lies), controlled topic variation, and interpretable linguistic analysis. Limitations include small corpus size (300 statements per topic), focus on written English opinions rather than news or social media, and reliance on feature engineering rather than learned representations. The work demonstrates that linguistic patterns do distinguish truth from lies, though with room for improvement. Follow-up work could explore larger corpora, different domains (news, social media, dialogue), multimodal features, and neural approaches. The dominance score metric for class-based analysis is a useful interpretability technique.