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

Summary

The paper asks whether written lies can be told apart from truthful statements automatically, and which linguistic signals mark deception. No labeled corpus existed, so the authors built one using Amazon Mechanical Turk. There are three topics: abortion, death penalty and best friend. For each topic, contributors wrote one statement that reflected their real view and one that argued the opposite. For the best-friend topic, the false statement described a disliked person as if they were a close friend.

Naïve Bayes and SVM classifiers were trained on lightly preprocessed word features. Within a topic, they reached about 70% accuracy against a 50% baseline. Cross-topic accuracy was lower but still above chance.

The authors also introduced a "dominance" score that compares how often LIWC word classes appear in deceptive versus truthful text. Deceptive texts lean toward references to other people and certainty words. Truthful texts lean toward self-reference, optimism and insight words. The work is an early demonstration that deception detection from text alone is feasible.

Key points

• Data: 100 true and 100 false statements per topic, averaging about 85 words each.
  • Every entry was checked by hand.
  • Two entries were dropped because their true and false versions were identical.
• Setup: tokenization and stemming only. Stopwords were kept and no feature selection was done. Evaluation used ten-fold cross-validation.
• In-topic accuracy:
  • Naïve Bayes averaged 70.8% and SVM averaged 70.1%.
  • The best-friend topic scored highest (75.0% NB, 77.0% SVM).
• Cross-topic accuracy (train on two topics, test on the third): about 59.8% for NB and 57.8% for SVM.
• LIWC dominance analysis (2001 lexicon, ~2,200 words in ~70 categories):
  • Classes most typical of lies: METAPH (1.71), YOU, OTHER, HUMANS and CERTAIN.
  • Classes most typical of truth: OPTIM, I, FRIENDS, SELF and INSIGHT (scores around 0.57–0.65).

Interesting ideas

• I1. Paired elicitation by the same author: each contributor writes both a truthful and a deceptive text on the same topic. This keeps author style and topic roughly constant, so the label difference is easier to isolate. The same design could help build controlled corpora for stance, sarcasm or persuasion.
• I2. Class-level dominance ratio: the score divides a word class's share of tokens in one corpus by its share in another. It gives an interpretable view of which lexical categories separate two text collections. It could be reused with any lexicon, such as sentiment, register or dialect word lists, including for Ukrainian resources.
• I3. Leave-one-topic-out evaluation: training on some topics and testing on an unseen one checks whether a classifier learns the target phenomenon rather than topic-specific vocabulary. This is a cheap sanity check for most text classification tasks.
• I4. Learning curves as a data-collection guide: plotting accuracy against the fraction of training data shows whether gathering more annotations is likely to pay off. This helps decide where to spend crowdsourcing budget.
• I5. Deliberately minimal baseline: using no feature engineering and no stopword removal shows how much signal raw word features carry on their own. Function words, which carry stylistic cues, stay in the feature set.