Using linguistically defined specific details to detect deception across domains

Abstract

Current automatic deception detection approaches tend to rely on cues that are based either on specific lexical items or on linguistically abstract features that are not necessarily motivated by the psychology of deception. Notably, while approaches relying on such features can do well when the content domain is similar for training and testing, they suffer when content changes occur. We investigate new linguistically defined features that aim to capture specific details, a psychologically motivated aspect of truthful versus deceptive language that may be diagnostic across content domains. To ascertain the potential utility of these features, we evaluate them on data sets representing a broad sample of deceptive language, including hotel reviews, opinions about emotionally charged topics, and answers to job interview questions. We additionally evaluate these features as part of a deception detection classifier. We find that these linguistically defined specific detail features are most useful for cross-domain deception detection when the training data differ significantly in content from the test data, and particularly benefit classification accuracy on deceptive documents. We discuss implications of our results for general-purpose approaches to deception detection.

Masking domain-specific information for cross-domain deception detection

“…On the other hand, if it is found to be domain-specific (e.g., quotes, parentheses, or compound terms like and/or), it is replaced by the symbol @. Furthermore, to consider all the numeric details usually given by truthful communicators (Vogler and Pearl, 2018), we mask numerals (e.g., one, two, three, etc.) with a single symbol +.…”

Building and benchmarking the motivated deception corpus: Improving the quality of deceptive text through gaming

“…One of the earliest is the work done by (Newman, Pennebaker, Berry, & Richards, 2003), which investigated linguistic clues to determine differences between truthful and deceptive statements. Vogler and Pearl (2019) used a support vector machine operating on linguistically defined features to classify the Ott corpus. They were able to achieve an accuracy of 87% using this method.…”

“…Humans, in fact are extremely poor even at identifying if a review is generated by a human or artificial intelligence (Hovy, 2016). This is in stark contrast to other linguistic tasks such as sentiment analysis (e.g., identifying if a text sample is praising or condemning something) where humans perform extremely well (Vogler & Pearl, 2019). Part of the purpose of this corpus is to provide an opportunity to evaluate human performance on deceptive text in an environment where they will be motivated to perform well.…”

“…However, in this respect, deceptive text has an additional complication that is not present in, for example, sentiment analysis. Sentiment in text is comparatively easy for humans to identify (Vogler & Pearl, 2019) and generate, and good samples are relatively simple to gather. Quality samples of deceptive text, on the other hand, are much more difficult to assemble.…”

Detecting Deception, Partisan, and Social Biases

“…quotes, parentheses, or compound terms like and/or ), it is replaced by the symbol @. Furthermore, to consider all the numeric details usually given by truthful communicators [209], we mask numerals (e.g. one, two, three, etc.)…”