How analysis of mobile app reviews problematises linguistic approaches to internet troll detection

State-sponsored internet trolls repeat themselves in a unique way. They have a small number of messages to convey but they have to do it multiple times. Understandably, they are afraid of being repetitive because that will inevitably lead to their identification as trolls. Hence, their only possible strategy is to keep diluting their target message with ever-changing filler words. That is exactly what makes them so susceptible to automatic detection. One serious challenge to this promising approach is posed by the fact that the same troll-like effect may arise as a result of collaborative repatterning that is not indicative of any malevolent practices in online communication. The current study addresses this issue by analysing more than 180,000 app reviews written in English and Russian and verifying the obtained results in the experimental setting where participants were asked to describe the same picture in two experimental conditions. The main finding of the study is that both observational and experimental samples became less troll-like as the time distance between their elements increased. Their ‘troll coefficient’ calculated as the ratio of the proportion of repeated content words among all content words to the proportion of repeated content word pairs among all content word pairs was found to be a function of time distance between separate individual contributions. These findings definitely render the task of developing efficient linguistic algorithms for internet troll detection more complicated. However, the problem can be alleviated by our ability to predict what the value of the troll coefficient of a certain group of texts would be if it depended solely on these texts’ creation time.