Mindful Tweet Hub

In the era of social media, mental health challenges like anxiety and depression are on the rise. This project first categorizes tweets into 10 mental health conditions: Eating Disorder, Schizophrenia, OCD, PTSD, Anxiety, Bipolar, Autism, Depression, ADHD, and Control using BERT classification. It then finds similar tweets within each category and presents them to users, fostering a sense of community and letting them know they're not alone in their struggles.

• Created BERT embeddings from the model that we fine-tuned using BertForSequenceClassification
• For all the tweets by removing the classification layer and getting the output bert embedding of CLS for each tweet.
• These embeddings were used to create HNSWFlat indexes from FAISS . Later, the embeddings were added to the clusters.
• Whenever the user enters a tweet, we calculate the BERT embedding for that user and search for the top 20 closest tweets in the HNSW. (It returns the distance, which we are using to get the similarity)
• Using these tweets we are finding the Okapi BM25 score for each tweet with the query tweet.
• The combined score of each embedding with these 20 tweets would be 0.8* (similarity from HNSW bert embedding matching) + 0.2"(bm25 similarity)
• We now get the top similar tweets, we would use the unique classes in these tweets to do a zero shot classification to get the maximum probable class in which the tweet can fall into.