Performance Evaluation of Various Embedding Techniques for Identification of Depressive Contents in Social Media

Depression has emerged as the most common mental disorder in today's world, causing experts to concentrate on early detection in order to minimize its long-term consequences, including the likelihood of suicidal thoughts leading to suicide. As societal dependence on social media as an essential means of interaction expands, it has evolved into a significant platform for individuals to interact and engage with others, surpassing conventional interactions in person. Understanding this transition, researchers have shifted their focus to using social media data to detect indicators of depression. There is a rising opportunity of identifying and tackling depression more effectively, potentially preventing its damaging repercussions, by using the large volume of information provided on these platforms. To provide a solution, we propose a model that can effectively identify depression in social media data. Considering three different embedding techniques such as TF-IDF, Word2Vec, and BERT model classified using conventional machine learning algorithms, we compare the performance of each embedding technique on two different datasets. Context-aware sentence embedding method BERT has achieved better accuracy compared to hand-crafted feature extraction and static word embedding methods.