Reconstructing truthful responses in psychological tests using NLP state-of-the-art models Relatore: Professore Giuseppe Sartori Corelatore : Professore Mauro Conti

Psychological questionnaires are a powerful tool to assess psychological conditions such as personality traits, or mental diseases such as depression. They are lists of sentences describing psychological symptoms (e.g. I always think about suicide). The major weakness of this tool is that in many of the scenarios where those questionnaires are used, the subjects who have to respond would benefit from showing a particular condition. Hence encouraging the respondees to answer not according to what they think the truth is, but according to what they think would make them appear the way they believe to be the best in the specific scenario they are into. So to tackle this problem, the first step is to understand when a response is deceiving and the consequent step is to reconstruct what would have been the truthful response. While several machine learning techniques have been successfully applied to detect dishonest responses, no approach I am aware of has been shown to handle the reconstruction of truthful responses properly. Furthermore, all the approaches tried so far both for the first and the second task required a specific model for each type of questionnaire. In this work, I show how can NLP state-of-the-art models be used in a transfer learning framework to address every questionnaire with a unique model, both in lie detection and response reconstruction, obtaining satisfactory results. In this work is also discussed how to improve this approach and further works that can possibly be done.