Linguistic Indicators of Depressive Symptoms in Conversations with Virtual Humans.

Depression is a major public health concern that currently lacks reliable biomarkers for diagnosis and assessment. This study aims to analyse linguistic patterns associated with depressive symptoms during interactions with Virtual Humans (VH). 40 participants were equally divided into a control group and a group exhibiting depressive symptoms. Each participant engaged in six semi-guided conversations with VHs designed to elicit basic emotions and powered by a Large Language Model to generate automatic responses. Participants’ speech was transcribed and linguistic features were extracted using LIWC. Non-parametric statistical tests were employed to examine differences between groups. Participants with depressive symptoms used more words associated with negation and exclusion, and referred more to negative emotions. Control participants based their speech more on certainty and temporal reference, and used more frequently words related to friendships and health. The results support the potential of VHs and linguistic patterns as valid biomarkers for assessing depression.