An analytical system for user emotion extraction, mental state modeling, and rating

Abstract Emotion is a crucial indicator of the mental state influenced by environmental cognition, health, and intention. Therefore, modeling emotional patterns from daily communications is useful and vital for behavioral studies. This paper presents an analytical system that combines psycho-linguistics, statistical modeling, and computational algorithms to extract, analyze, and rank users’ emotions from their emails. A new category of twelve emotions is proposed to identify a balanced set of six positive emotions and their negative counterparts, respectively. A lexicon-based algorithm is proposed to extract emotions from individual emails at the sentence level by combining thesaurus and parts-of-speech. A Markov chain model is then proposed for the time-dependent emotion stochastic process. Statistical characterization of the Markov chain model provides insights from four aspects: repetition, revisits, steady states, and emotional stability. Finally, an algorithm is discussed to compute the relative ratings of user emotions and subsequently ranking them from the most emotional to the least. As the verification that the proposed system is able to distinguish detailed emotional behaviors caused by mental activities, the Enron email dataset is used to model emotional patterns from 22 executives who were deemed guilty in the court of law vs. 21 non-guilty. Our analysis demonstrates that the proposed system uncovers emotional patterns that are more distinct compared to the traditional emotion categorizations. It is also observed that the guilty subjects express more positive emotions in their emails compared to their non-guilty counterparts, while they are not more joyous than the non-guilty. Such insights have implications for employee risk assessment in organizations. The proposed approach can also be applied to other related areas such as mental health.