SCSE20-0985
AI for Healthcare: Analysis on Depressive Social Media Texts
Mental health has been an increasingly challenging issue to tackle in this era due to the stressful environment we are living in. One such example of mental health illness is depression. Depression is a mood disorder described as feelings of sadness, loss or anger that interfere with one’s everyday activities. Depression has existed as a problem in this society for many years.
With technological advancements, social media platforms serve as a place for depressed personnel to seek help, hoping to feel better in one way or another. However, their problems are often neglected by others on the internet. If not detected quickly and accurately, one’s depression may develop into more serious issues such as suicidal thoughts.
Research has shown that nearly 300 million people in the world suffer from depression every year. Measures to assess depression include clinical judgement or structured interviews, but a more common method is the use of social media analysis. Social media helps to detect depression by analysing posts on social media platforms. This method is preferred as expressing one’s feelings online has become the new norm, and processing of social media data can take place quickly, so authorities are able to intervene at an earlier stage.
This project thus aims to analyse depressive texts from social media such as Twitter and Reddit by building various deep learning models for the different main tasks, hoping that we can detect depression and the cause of depression at an earlier stage. These main tasks include Classification, Emotion Intensity Prediction and Emotion-Cause Pair Extraction.
The focus of this project will be to leverage on the use of Deep Learning and NLP to do an in-depth analysis on depressive social media contents. Specifically, we will begin with the more common task of predicting whether a particular text contains sentiment of depression. Then we will dive deeper to create a depression metric and predict depression scores for all of the data that are labelled as depressive content. Lastly, we will experiment with the more uncommon task of Emotion-cause Pair Extraction, which is to determine how accurate the model can extract the emotion and the cause of depression from text clauses.
In this project, there will be a total of three main tasks and one subtask to perform an in-depth analysis on depressive social media texts. These tasks are stated below:
Main Tasks:
- Emotion Classification
- Emotion Intensity Prediction
- Emotion-cause Pair Extraction
Subtask:
- Text Summarisation (Extractive & Abstractive)
The whole project will be coded in Python 3. Both the Jupyter Notebook and the Google Colaboratory will be used to implement the depression analysis on social media data. Codes written in Jupyter notebooks are mainly dealing with importing of data and data cleaning for later analysis. Codes written in Google Colaboratory are mainly dealing with training of models for the different main and subtasks. Google Colaboratory is preferred for training of models because it has a free Graphics Processing Unit (GPU) that can be used to speed up training.