DECEPTION DETECTION ON LIES, REVIEWS AND TRIALS

Abstract

Since past few years, deceptive contents such as fake reviews posted on online shopping sites, also identified as opinion spam, or deceptive contents in any form either verbal or textual have become a nuisance because of exponential advancement in information technology and communication and hence the ease of creation/distribution of any kind of information.

Fake reviews affect the consumers‟ decision making abilities and the reputation of stores across all the platforms. Also, the need for tackling and identifying deceptive contents from the day-to-day life using the headway in computation and technology is also being understood. The problem of opinion spamming was discovered not long ago, even then it began to be a promising research front because of the ever-growing abundance of computer generated data, formally, text based computer mediated communication (TB CMC). Since it has become fairly easy to write and/or propagate any deceptive contents with just a click, so the algorithms tackling such problems need to out-perform them to eradicate them or at least stop them as quickly as they are generated.

The lack of efficient ideas and algorithms, along with the technology to implement them pose a huge hurdle in the way of automating the task of deception detection, since human

v

experts can‟t be relied for it anymore, due to various factors such as lack of time, efficiency, knowledge, manpower, great amount of data generated, ever changing forms of deceptive data etc.

In this project, we have developed linguistic, syntactic and semantic models to detect fake/deceptive contents especially fake reviews, lies and deceptive speeches. We experimented using different methods of feature extraction for different categories of features and combined and evaluated them on various classification models of machine learning. The evaluation was done on open domain deception data, product reviews data and real life trial data to further identify the features and classification techniques that suit the domain and purpose of the experiment.