Detecting Distressed and Non-distressed Affect States in Short Forum Texts

摘要

Improving mental wellness with preventive measures can help people at risk of experiencing mental health conditions such as depression or post-traumatic stress disorder. We describe an encouraging study on how automatic analysis of short written texts based on relevant linguistic text features can be used to identify whether the authors of such texts are experiencing distress. Such a computational model can be useful in developing an early warning system able to analyze writing samples for signs of mental distress. This could serve as a red flag, signaling when someone might need a professional assessment by a clinician. This paper reports on classification of distressed and non-distressed short, written excerpts from relevant web forums, using features automatically extracted from input text. Varying the value of k in k-fold cross-validation shows that both coarse-grained and fine-grained automatic classification of affect states are generally 20% more accurate in detecting affect state than randomly assigning a distress label to a text. The study also compares the importance of bundled linguistic super-factors with a 2k factorial model. Analyzing the importance of different linguistic features for this task indicates main effects of affect word list matches, pronouns, and parts of speech in the predictive model. Excerpt length contributed to interaction effects.