BERT-Based Approach for Suicide and Depression Identification

Abstract

With the advent of the Internet, there is an increasing requirement for sophisticated and intelligent systems that can operate efficiently handle the identification on social media about health-related issues, such as depression and suicide recognition. The data generated by social media users is unstructured and unreliable. The text representation and deep learning algorithms used, however, give only a limited amount of information’s as well as expertise regarding the various user-supplied texts. We started with bidirectional long short-term memory (Bi-LSTM) with progressed toward transformer models similar as bidirectional encoder representations from transformers (BERT). In our research, we discovered that they perform considerably superior to traditional deep learning architectures such as Bi-LSTM and BERT. We have designed our own data collecting platform using Reddit, one of the majority popular social networking sites. Finally, we have used BERT plus Bi-LSTM to efficiently evaluate as well as notice indicators of depression plus suicide in social media posts.