Deceptive Opinion Spam Detection

• Explores deceptive opinion spam, which are fictitious opinions deliberately written to sound authentic.
• Integrates psychology and computational linguistics to detect opinion spam.
• Develops a classifier with nearly 90% accuracy on the gold-standard opinion spam dataset.
• Reveals a relationship between deceptive opinions and imaginative writing through feature analysis.

Introduction

• Review websites with user-generated opinions have increasing potential for monetary gain through opinion spam.
• Opinion spam can range from self-promotion to deliberate review fraud.
• Focuses on deceptive opinion spam, which are fictitious opinions written to deceive the reader.

Related Work

• Focuses on opinion spam detection and identifies the challenges around this task.
• Opinion spam is both widespread and different in nature from either e-mail or Web spam.
• Aims to find deceptive opinion spam detection beyond what humans can easily recognize.

Dataset Construction and Human Performance

• Details the process of gathering and validating deceptive and truthful hotel reviews.
• Utilizes Amazon Mechanical Turk (AMT) to solicit deceptive opinions.
• Pays Turkers one US dollar for an accepted submission, asking them to write a fake positive review for a hotel.
• Mines all 5-star truthful reviews from the 20 most popular hotels on TripAdvisor in the Chicago area.
• Assesses human deception detection performance and determines that human judges are not effective at this task.

Automated Approaches to Deceptive Opinion Spam Detection

• Explores three automated approaches to detect deceptive opinion spam: genre identification, psycholinguistic deception detection, and text categorization.
• Assesses whether truthful and deceptive reviews have a relationship with part-of-speech (POS) tags in text
• Uses Linguistic Inquiry and Word Count (LIWC) in aim to detect personality.
• Uses three n-gram feature sets, like unigrams, bigrams, and trigrams when text categorizing and contrasting detection of different texts.

Classifiers

• Uses Naive Bayes and Support Vector Machine classifiers.
• Employs the SRI Language Modeling Toolkit (Stolcke, 2002) to estimate language models for truthful and deceptive opinions.

Results and Discussion

• Evaluates deception detection, discovering automated classifiers outperform human judges in nearly every metric.
• Human judges use unreliable cues for deception, as supported by one study of online dating.
• Psycholinguistic approach performs more effectively than the genre identification method.
• Suggests that truthful opinions use concrete language and focuses on elements external to the hotel.
• Identifies a plausible relationship between deceptive opinion spam and imaginative writing linked to POS distributional similarities.