Inspiration
Several recent surveys of students suggest their mental well-being has been devastated by the pandemic’s social and economic consequences, as well as the continued uncertainty about college education and post college careers. Existing and prevalent problems in colleges' mental health support systems, such as long wait times and under-staffing, accompanied by increased demand leave students without the help they need. These problems must be fixed by academic institutions. However, these institutions cannot help students if students do not know where to go.
What it does
ALECC, which stands for Aggies Love Existential Crisis Chatbots, is an SMS chatbot. It can predict depression-related sentiments for either 1) the user's response to the chatbot over SMS or 2) public tweets, given a twitter handle. It generates these predictions using a trained machine learning model.
How we built it
Building the chatbot*: We used the Twilio API to create a chatbot that can be accessed by texting a phone number. Our project was coded in python, so we used the Flask framework in order to run our project. The phone number was connected to our Flask app using a ngrok web hook.
Gathering data: We used the Twint library to pull tweets directly from Twitter. Our total dataset contained about 10, 500 tweets, which was a combination of a pre-labeled tweet dataset as well as a collection of our own tweet dataset that we manually classified. These tweets were then processed by removing handles, links, special characters, and then tokenized as lists of words.
Creating a classifier: We used a Sklearn Naive Bayes model to create a classifier that would categorize whether or not a tweet was a potential mental health concern. We had a 70/30 test/train split, and our tweets were vectorized using TF-IDF (term frequency–inverse document frequency) before feeding them to our model. Based on our testing data, we obtained 0.88 accuracy and a 0.75 F1 score. Lastly, we added a list of red flag keywords to address the case of potential false negative predictions.
Challenges we ran into
We ran into the issue of our training data either coding everything as either having a mental health concern or not a mental health concern - there seemed to not be a middle ground.
Take a look for yourself
Text "hello" to 530-298-8024 to chat with ALECC!
Log in or sign up for Devpost to join the conversation.