RatioLogo
Back

Unmasking Lies by Counting Suffixes

What if the most effective way to catch a lie isn't by fact-checking the claims, but by counting the suffixes?

In the high-stakes arms race against digital misinformation, most tools are built for English. This leaves agglutinative languages—those like Turkish, Finnish, or Hungarian that build complex meanings by stacking suffixes onto word roots—vulnerable to a flood of undetected fabrications. Now, a breakthrough study from Hacettepe University has produced FaNLexTR, the first dedicated lexicon designed to decode the specific linguistic "DNA" of Turkish fake news.

The research matters because it moves beyond the slow, resource-heavy deep learning models that often struggle with the sheer morphological complexity of Turkish.

The Study's Core Findings

By analyzing a massive corpus of 84,734 unique articles, researchers discovered that the "signal" of a fake story is often hidden in the grammar itself.

Distinctive Linguistic Markers

The team found that deceptive content in Turkish is statistically more likely to:

  • Use informal slang.
  • Address the reader directly with specific verb endings.
  • Contain frequent misspellings.

In fact, fake news from sources like teyit.org showed a misspelling rate of 0.396 per sentence, drastically higher than credible news.

Computational Model Performance

To capture these patterns, the researchers tested four different computational models. The results were striking:

1. The RAW Form Model

This approach, which looks at words exactly as they are written—including typos—emerged as the most effective.

  • Recall: 0.929
  • Meaning: It caught nearly 93% of the fake news it encountered.

2. The SUFFIX Model

Remarkably, this model ignored the actual meaning of words entirely, looking only at the strings of suffixes attached to them. Even without "knowing" what the news was about, it achieved strong performance.

  • Accuracy: 0.864

3. The ROOT/STEM Model

This version focused on the semantic core of words, proving that even at the root level, deceptive language carries a distinct mathematical signature.

  • Accuracy: 0.807

Key Implications and Limitations

The study confirms a global phenomenon: falsehoods are often "simpler" and more aggressive than the truth. In Turkish, this manifests as shorter articles that use slop—linguistic shortcuts and informalities—to provoke emotional responses.

Current Research Hurdles

The path to a perfectly clean feed still has hurdles:

  • Dataset Imbalance: The study utilized 1,188 fake articles against 83,546 valid ones.
  • Morphological Ambiguity: The Turkish language's complex word structures can occasionally trip up NLP engines, leading to minor misclassifications.

While currently optimized for text-only analysis, the FaNLexTR workflow provides a scalable blueprint for linguistic defense in any suffix-heavy language, offering a faster, more agile shield against the spread of misinformation.

Reference: Mertoğlu, U., & Genç, B. (2020). Lexicon generation for detecting fake news. Department of Computer Engineering, Hacettepe University.