Master of Science in Artificial Intelligence, Queen Mary, University of London, 2020-2021 Dissertation Code.
The advent of social media has enabled new forms of socialisation which have become increasingly essential to the daily lives of millions of people. Many people, using various social media platforms have been experienced an unprecedented level of connectivity. Social Networking sites have gained incredible prominence with Facebook registering 2,740,000,000 million users as of January 2021. (statista, 2021) Social media is fundamentally an interactive technology which allows users to create and share information. This user- generated content primarily consists of text-based posts and/or comments, possibly accompanied with digital photos or videos, and the corresponding metadata. Such broad use has enabled increased communication, allowing people to stay up to date with family and friends, join and promote social causes, and consume content of textual and graphic nature. Social media has also facilitated damaging forms of interaction such as cyber-texting, sexting, and online stalking. Notably, excessive use of these online platforms has been shown to fuel feelings of Anxiety, Depression, isolation, and FOMO (fear of missing out). (Obar & Wildman, 2015). Mental health awareness has increased significantly in recent years, owing its origins to the mental hygiene movement, initiated in 1908, which was driven by a desire to improve the treatment and quality of life of people with mental disorders. (Bertolote, 2008) With approximately 1 in 4 people in the UK suffering from a mental condition every year, and 1 in 6 suffering from a common disorder such as Anxiety or Depression, such issues have become widespread and prevalent. It has been noted that people are commonly readily eager to express their views anonymously online rather than in person. (Al-Saggar & Nielsen, 2014) Using various applications, anonymous users are likely to discuss their mental health problems on an online platform. (Hanwen Shen & Rudzicz, 2017) This is a positive trend due to patterns of hidden behaviours exhibited by people when fearing stigmatisation, i.e. people have been more likely to under report mental health issues compared to other health conditions due to the associated stigma. (Bharadwaj & Suziedelyte, 2017) Therefore, the anonymisation of identity, allowed by online interaction, has allowed people to discuss their mental issues without risking social stigma. By discussing their issues openly and without fear of stigmatisation, people have been less likely to under-report and therefore, have been giving more accurate accounts of potential symptoms. People may also engage in discussions but be unaware or not willing to consider that they are potentially suffering from a mental health condition. It is unrealistic to expect mental health professionals to inspect and review large and ever-increasing amounts of data. An automated unit for the purposes of mental health text classification could alert online users if the text they are posting would be indicative of a mental health disorder and would encourage them to seek professional help. This project focused on using the social media generated textual data to train a Deep Learning based model with the aim of detecting whether the person posting on social media is likely to be suffering from a mental health disorder based on the content they are posting onto social media. Such a task would fall into the field of Natural Language Processing (NLP). Core NLP techniques were traditionally dominated by machine learning methods using linear methods such as support vector machines or linear regression, trained over very high dimensional yet sparse feature vectors. The field has since found increased success in recent years by making use of non-linear neural network models over dense inputs, a technique known as Neural Networks. (Goldberg, 2015) Such a technique, specifically when extended to Deep Learning, allows computational models that are composed of multiple processing layers to develop representations of data with various levels of abstraction. Deep Learning models have dramatically improved the state-of-the-art in various domains of application. The strength of these models is their capability to discover intricate structure and nuanced patterns in large datasets. This is accomplished by using the backpropagation algorithm to indicate how a machine should change its learnable weights which are used to compute the representation in each layer from the representation in theprevious layer. (LeCun, et al., 2015) The application of Deep Learning to Natural Language Processing tasks yields several advantages: superior performance at pattern recognition tasks, and the capability of end-to-end training (little or no domain knowledge is needed prior to the system construction). Deep Learning however is a data hungry process and is hence not suitable for small quantities of data. Its resultant models are typically black box, making them difficult to understand due to the continuing lack of theoretical foundation. Furthermore, the cost of training Deep Learning models is computationally expensive. (Li, 2018) This work relies on the notion that the text posted by a user when suffering from a mental health condition would contain different, detectable features when compared to the text generated by the same user when not suffering from a mental health condition.
This section includes the files necessary to train the mental health classifiers.
This section includes the files used to deployment a Gradio based web demo and a terminal based demo using flask.
This section includes deliverables including the Dissertation paper discussing the project, a reflective essay focusing more on ethics, and potential social impacts of such a project, and finally a power point presentation briefly discussing an overview of the project.