Care2Share
2020 has been a strange year to say the least, largely dominated by Coronavirus. When the pandemic reached my doorstep, everything changed. All the plans I had , suddenly felt like and most probably still are impossible for the next couple of years. Those few months in Isolation were rough , thankfully though everyone in my family was lucky to survive. After the situation got better and I analyzed how I lived my life during the previous months, one thing that was immediately obvious was how I completely isolated myself from everyone. I almost completely stopped all social interaction which after more research I found was the wrong thing to do when one is going through tough times. I was fortunate to have the support structure around me to get through this time but I realize that many people might not be as fortunate as I was. Thus for my university capstone project, I decided to tackle this issue.
My main idea was to create a platform/environment in the form of a mobile application which users can feel comfortable to use when they are not feeling mentally well. The app will have 3 main components , an option to chat and share their thoughts anonymously with other users on the app, use a chatbot designed for any issues they might be facing related to mental health and use a mood diary to keep track of their moods.
Since this was the first time I was building an app, I knew that there was going to be a steep learning curve. I chose flutter as the framework for building the app as it was platform independent and open source . The name I decided for the app was care2share, to symbolize that as a community we can create a space for everyone to improve their mental well being by being polite and caring for each other’s thoughts and feelings.
Initially, a conversational chatbot was planned and designed using DialoGPT, a state of the art response generation model built using OpenAI’s GPT-2 for users who did not feel comfortable interacting with other users on the app. DialoGPT was trained using approximately 147M multi-turn dialogue from reddit discussion threads. After fine-tuning the model with a dataset collected from counsel-chat where therapists respond to user queries related to issues they are facing, the model was deemed unsafe to use for the app due to the sarcastic and rude responses it sometimes generated. These responses are generated most likely because the data used to train the DialoGPT might have contained similar responses. Each user might react differently to such responses when they are not feeling mentally well, potentially some users might feel worse after seeing them.
While more research and experimentation is done to fix the problems with DialoGPT ,in the mean time to replace it a Q&A type Bot was made using the data collected from counsel-chat website. User will select a category he/she needs help with and subsequent responses would be provided. The data collected from counsel chat are real therapists’ responses so there is no issue of responses having negative impact on users.
The mood diary used was inspired by this blog.
To detect depression and suicidal thoughts from text messages and provide recommendations in the form of blogs/articles/music via mail on a weekly basis , a zero shot learning model from HugginFace is utilized. The model will also intervene if it detects suicidal thoughts or hate speech in the messages.
The app is not perfect and could use some improvements but overall I am proud of myself for having an attempt at improving mental well being of people who might be in a similar spot as I was.
References
- https://counselchat.com/
- https://discuss.huggingface.co/t/new-pipeline-for-zero-shot-text-classification/681
- https://medium.com/flutter-community/develop-a-mood-diary-flutter-app-in-2-hours-ca54f8bf6703
- https://towardsdatascience.com/counsel-chat-bootstrapping-high-quality-therapy-data-971b419f33da
- https://github.com/microsoft/DialoGPT
