Accepted for/Published in: JMIR Formative Research
Date Submitted: Jan 2, 2022
Date Accepted: Jul 20, 2022
Warning: This is an author submission that is not peer-reviewed or edited. Preprints - unless they show as "accepted" - should not be relied on to guide clinical practice or health-related behavior and should not be reported in news media as established information.
Multilevel Depression Detection from Bengali Text Data: A Study Using Natural Language Processing Techniques
ABSTRACT
Background:
There are a myriad of language cues that indicate depression in written texts and Natural Language Processing researchers have proven the ability of machine learning and deep learning approaches to detect these cues. However, till date these approaches bridging NLP and domain of mental health for Bengali literature are not comprehensive. The Bengali speaking population can express emotions in their native language in greater detail. Therefore, we constructed a procedure for extracting textual information from Bengali literature for predictive analysis with the aim of detecting the degree of depression.
Objective:
Our goal is to detect the severity of depression from Bengali texts.
Methods:
We conducted a study using Bengali text-based data from blogs and open source platforms. In this study we developed our own structured dataset and designed a labeling scheme with the help of mental health professionals and adhering to DSM-5 during the process. We employed 5 machine learning models: Kernel Support Vector Machine, Random Forest, Logistic Regression K-Nearest Neighbor, Complement Naive Bayes for detecting the severity of depression. For the deep learning approach we used Long Short Term Memory Units and Gated Recurrent Units coupled with convolutional blocks or self-attention layers. Finally, we aimed for enhanced outcomes by utilizing state-of-the-art pre-trained language models.
Results:
The independent Recurrent Neural Network models have yielded the highest accuracies and weighted F1-scores. Gated Recurrent Units in particular produced 80% accuracy. The hybrid architectures could not surpass the Recurrent Neural Networks in terms of performance. Kernel SVM with bag-of-words embeddings generated 77% accuracy on test data. We employed validation and training loss curves to observe and report the performance of our architectures. Overall the number of available data remains the limitation of our experiment.
Conclusions:
The findings from our experimental set-up indicate that machine learning and deep learning models are fairly capable of assessing the severity of mental health issues from texts. For the future, we suggest more research endeavors to increase the volume of Bengali text data in particular so that modern architectures reach improved generalization capability.
Citation
