Automatic detection of deceptive opinions using automatically identified specific details

Distinguishing deceptive opinions — that is, fabricated views disguised to be genuine — from honest opinions is a hard problem. Deceptive opinions can include things like the false expression of a controversial opinion, a misleading review of an item or service bought online, or deceitful interviews. Unlike many tasks involving language, detecting deceptive opinions through text alone turns out to be quite difficult for humans. Ott et al. (2011) demonstrated this by asking a group of students to judge whether 40 online reviews were truthful or deceptive. These reviews were drawn from their Deceptive Opinion Spam Corpus, introduced in the same paper, and so included an equal number of truthful and deceptive reviews. The students performed at roughly 60% accuracy — only slightly better than the baseline, chance performance of selecting one of two choices (truthful or deceptive), which is 50%. Notably, the students were psychologically biased towards judging more opinions as truthful rather than deceptive. This poor performance suggests that detection of deceptive opinions is a complex area that can greatly benefit from unbiased computational analysis. Much of the research performed on deceptive opinions has used online reviews from Ott et al. (2011)’s corpus as a benchmark because these data are a rich source of opinion spam, a type of deceptive opinion. As e-commerce burgeons, online reviews are becoming increasingly important to company reputations and consumer product assessment. Due to its influential impact on potential customers, deceptive opinion spam is being produced to deceive potential consumers. Whether the opinion spam is sponsored by a company wanting to promote its services (positive opinion spam) or a business maligning its rival with false claims (negative opinion spam), accurate