Detecting Deceptive Opinions: Intra and Cross-Domain Classification Using an Efficient Representation

Abstract

Online opinions play an important role for customers and companies because of the increasing use they do to make purchase and business decisions. A consequence of that is the growing tendency to post fake reviews in order to change purchase decisions and opinions about products and services. Therefore, it is really important to filter out deceptive comments from the retrieved opinions. In this paper we propose the character n-grams in tokens, an efficient and effective variant of the traditional character n-grams model, which we use to obtain a low dimensionality representation of opinions. A Support Vector Machines classifier was used to evaluate our proposal on available corpora with reviews of hotels, doctors and restaurants. In order to study the performance of our model, we make experiments with intra and cross-domain cases. The aim of the latter experiment is to evaluate our approach in a realistic cross-domain scenario where deceptive opinions are available in a domain but not in another one. After comparing our method with state-of-the-art ones we may conclude that using character n-grams in tokens allows to obtain competitive results with a low dimensionality representation.

Applying automatic text-based detection of deceptive language to police reports: Extracting behavioral patterns from a multi-step classification model to understand how we lie to the police

“…To the best of the authors' knowledge, to date the best designed technique in a similar domain is by Ott et al [43] that accurately classify 89.8% deceptive hotel reviews by using a combination of bigrams, a psycholinguistic deception detection tool LIWC (Linguistic Inquiry and Word Count software [44]) and a SVM classifier. However, this papers' methodology exceeds current accuracy results regarding the automatic detection of verbal deception, as illustrated in Section 4.1, both in the same domain as Ott et al [43] , and in other domains [36,42,33,5,24,22]. Despite of the domain difference (i.e., false reports VS fake reviews) and the nature of the document (i.e., facts collected and written by a policeman after questioning the victim of a felony VS a first person written text), the insights drawn by analyzing the structure of the estimated probability model seem to lead to common patterns also identified by other researchers in the fake reviews domain.…”

“…Positive Sentiment Hotel Opinions This dataset, first introduced in [43] , and used in [43,42,24] and [5] , is comprised of 400 deceptive and 400 truthful positive reviews on 20 hotels in the Chicago area. VeriPol's results against the aforementioned state-of-the art approaches are shown in Table 2.…”

“…Positive and Negative Sentiment Hotel Opinions This dataset is first introduced in [42] and extends the Positive Sentiment Hotel Opinions dataset, by intro-ducing 400 deceptive and 400 truthful negative reviews on 20 hotels in the Chicago area. Some authors consider the positive and negative dataset separately [42,24,5] (see Table 3), while others jointly [42,5] (see Table 4).…”

“…Detecting automatically verbal deception [11], in court [12], in reviews [47,5,23] or in political debates [1] is a very hot topic in the research community. Most contributions in this line are mainly concerned with the domain of spam detection and fake reviews [43,32,22].…”

“…In summary, we are dealing with a task of detection and classification of texts that use verbal deception. This field, although currently a hot topic [11,1,5], is quite new which makes it difficult to compare our contributions against others concerned with the detection of false or deceptive texts [40,19,21,59]. These works, that are mostly centered on the domain of spam detection and fake reviews [43,32,22,47], make use of traditional ML techniques such as Naive Bayes, Support Vector Machine (SVM), logistic regression or decision trees [8,25].…”