Detecting Temporal Cognition in Text: Comparison of Judgements by Self, Expert and Machine

Background: There is a growing research focus on temporal cognition, due to its importance in memory and planning, and links with psychological wellbeing. Researchers are increasingly using diary studies, experience sampling and social media data to study temporal thought. However, it remains unclear whether such reports can be accurately interpreted for temporal orientation. In this study, temporal orientation judgements about text reports of thoughts were compared across human coding, automatic text mining, and participant self-report. Methods: 214 participants responded to randomly timed text message prompts, categorically reporting the temporal direction of their thoughts and describing the content of their thoughts, producing a corpus of 2505 brief (1-358, M = 43 characters) descriptions. Two researchers independently, blindly coded temporal orientation of the descriptions. Four approaches to automated coding used tense to establish temporal category for each description. Concordance between temporal orientation assessments by self-report, human coding, and automatic text mining was evaluated. Results: Human coding more closely matched self-reported coding than automated methods. Accuracy for human (79.93% correct) and automated (57.44% correct) coding was diminished when multiple guesses at ambiguous temporal categories (ties) were allowed in coding (reduction to 74.95% correct for human, 49.05% automated). Conclusion: Ambiguous tense poses a challenge for both human and automated coding protocols that attempt to infer temporal orientation from text describing momentary thought. While methods can be applied to minimize bias, this study demonstrates that researchers need to be wary about attributing temporal orientation to text-reported thought processes, and emphasize the importance of eliciting self-reported judgements.