Depression Severity Estimation on the InternetNew Models and Resources

Abstract

On the one hand, there is extensive evidence from medicine and psycholinguistics fields of changes in language usage from people suffering from mental health problems. On the other hand, social media platforms provide a vast repository of written language.There is a recent trend in computational linguistics where researchers aim to exploit social posts to detect individuals at risk. In this thesis, we follow that line in the field of depression detection. A shortcoming in actual research efforts is the need for more interpretability of the models’ decisions. To mitigate that problem, we investigate the development of models based on validated clinical symptoms to identify depressive signs. The contributions of this thesis are three-fold: (i) new models for depression severity estimation based on symptom markers, (ii) the creation of new datasets for helping the development of new symptom-based approaches, and (iii) the exploration of recent massive large language models for helping with the scaling up of the datasets construction. As a final step, we incorporate the above contributions into a demonstrative platform to be used by health professionals. This thesis contributes to advancing the understanding and detection of depression through symptom markers, and lays the foundation for future research in this critical area of depression detection on social media.