Depression Detection in Bengali Social-Media Text via Clinically Grounded Linguistic Feature Fusion and Interpretable Machine Learning
This paper presents a transparent, reproducible machine learning pipeline that achieves high accuracy (91.69% F1) in detecting depression in Bengali social media by fusing clinically grounded psycholinguistic markers with standard text features and employing an interpretable ensemble model.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Depression is a heavy burden that affects millions of people across the globe, often leaving them feeling isolated and unable to function in their daily lives. In many parts of the world, the number of trained mental health professionals is far too small to meet the need, creating a gap where suffering goes unnoticed. In Bangladesh, this shortage is particularly acute, with a single psychiatrist often responsible for hundreds of thousands of people. Because of this, many individuals turn to social media to express their pain, using their native language to share thoughts they might not say aloud to a stranger. For years, scientists have tried to build computer programs that can read these digital posts and spot the signs of depression, but most of these tools have been designed for English speakers. This leaves a vast population of Bengali speakers without access to automated screening, even though their language carries the same emotional weight and distress as any other.
A team of researchers from Daffodil International University has now built a new system designed specifically to understand Bengali social media text. Their goal was not just to create a program that guesses correctly, but to build one that a human doctor could trust and understand. They started with a massive collection of nearly 22,000 posts from Bengali social media, carefully labeled by humans as either showing signs of depression or serving as a normal control. The researchers knew that simply counting words was not enough, so they looked for specific patterns of thinking that psychologists have long associated with depression. They focused on five distinct ways the mind works when it is struggling: the use of absolute words like "always" or "never," a sense of hopelessness, a heavy focus on the self, a feeling that the pain will last forever, and the use of negation to deny reality. They also tracked words that directly express deep distress or thoughts of self-harm.
To make these patterns visible to the computer, the team did something clever. They did not just feed the raw text into the machine. Instead, they marked the text with special signals whenever they found one of these five thinking patterns. If a person wrote "I cannot do this," the computer was taught to see the word "cannot" as a single, special unit that carried extra meaning. At the same time, the system counted how often these patterns appeared in each post, turning the emotional tone into a simple number. They combined these emotional markers with standard ways of analyzing how words and letters are used, creating a rich picture of each post that included both the specific words chosen and the underlying psychological state.
The researchers then tested several different types of computer learning models to see which one could best use this information. They found that a combination of three different models worked better than any single one alone. One model was good at spotting simple patterns, another was excellent at handling complex relationships between words, and a third was very efficient at processing the data. By letting these three models vote on the final answer, with the most reliable ones having a louder voice, the system became highly accurate. When they tested this combined system on a set of posts it had never seen before, it correctly identified the presence of depression in more than 91 percent of the cases. It was also very good at avoiding false alarms, meaning it rarely labeled a normal post as depressed when it was not.
What makes this work stand out is not just its high accuracy, but its transparency. Unlike many modern computer systems that act like a black box, giving an answer without explaining why, this system can show its work. The researchers were able to point to the specific words and patterns that led the computer to its conclusion. For example, if a post was flagged as depressed, the system could highlight that it contained a high density of words related to hopelessness or self-focus. This ability to explain its reasoning is crucial for medical applications, where a doctor needs to understand the basis of a recommendation. The team also checked their work carefully and found that when the computer made a mistake, it was often because the original human label for that post was confusing or incorrect, rather than because the computer failed to understand the language.
The study confirms that the signs of depression are just as visible in Bengali text as they are in English, provided the right tools are used to look for them. The researchers did not rely on expensive, complex artificial intelligence that requires massive computing power. Instead, they used a method that is fast, affordable, and easy to audit. They demonstrated that by grounding their technology in established psychological theories, they could build a tool that is both powerful and trustworthy. While this system is designed as a screening aid to help identify people who might need support, and not as a tool to diagnose individuals on its own, it offers a promising path forward for mental health care in regions where professional help is scarce. It shows that with careful design, technology can be used to listen to the voices of those who are suffering, even in languages that have been overlooked for too long.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.