Opinion spam encompasses deceptive or fraudulent opinions—particularly reviews—posted online to manipulate reader perceptions. Unlike duplicate spam (verbatim copies) or off-topic spam (irrelevant content), opinion spam consists of authentic-sounding reviews that are deliberately written to be deceptive: either fabricated positive reviews to artificially inflate ratings, or negative reviews to damage competitor reputation. This category is particularly challenging because the text itself appears plausible and on-topic, but the intent and authenticity are compromised.
Key papers
- 2011-ott-deceptive-opinion-spam — Landmark dataset paper introducing 800 gold-standard hotel reviews (400 truthful, 400 deceptive); compares detection approaches using genre identification, psycholinguistic features (LIWC), and n-gram text categorization; finds combined LIWC+bigrams achieves 89.8% accuracy while humans perform at near-chance; reveals deceptive opinions correlate with markers of imaginative writing.
- 2019-adelani-sentiment-preserving-fake-reviews — Demonstrates high-quality fake review generation using fine-tuned GPT-2; shows both human and automated detectors struggle with generated reviews that preserve original sentiment.
Related topics
- fake-reviews (broader category including generation and detection)
- spam-detection (general framework)
- deception-detection (linguistic approaches to detecting lies)