Using Natural Language Processing (NLP), Sam is a bot designed to detect mental distress by scanning text content on Discord and Reddit. When less severe signs of distress are detected, Sam sends the user (individual experiencing distress) supportive text messages and offers mental health resources. In situations where Sam detects increased levels of distress, such as medium-to-high suicide risk, it notifies moderators to intervene and provides information to reach established sources, such as national helplines.
Created in Python, this bot was built using NLP models and datasets examining behavioral analysis in text and launched using Reddit and Discord APIs.
To identify persons in distress, we used Multinomial Naïve Bayes and Support Vector Machine models to analyze keyword data. We trained these supervised learning models on a vast dataset containing Wikipedia comments labeled by human raters for toxic behaviour.
In turn, we constructed a database of social media content using the following labels: ‘suicide’, 'lonely,' 'stressed,' 'normal,' 'non-suicide,' 'not depressed,' 'depressed,' 'anxious'. Multi-class text classification was then conducted using Decision Tree, Linear Support Vector, and Naive Bayes classifiers.
Through this project, we learned how to clean natural language data, process it using various machine learning models stated above, and then run these models through the Discord bot framework. While doing so, we faced a problem with the accuracy prediction dropping drastically as we include more labels for the text.
Finally, to maximize Sam’s capabilities, funding will be directed towards the following: Development: Detection of discriminatory content and false positives from “troll” behaviour in text and non-text posts. Enhanced support based upon user preferences and location, pending their consent. Human capital: Translation of text interventions into multiple languages. Collaboration with psychologists to develop and validate empathic scripts. Cross-platform integration: Embedment on more platforms, including WhatsApp and Telegram.
Log in or sign up for Devpost to join the conversation.