BDMentalText: A Bangla Text Dataset for Depression Detection and Mental Health Sentiment Classification (V1)

Abstract

BDMentalText V1 is a labeled Bangla (Bengali) text dataset for detecting depression-related language and classifying mental-health-related sentiment in a low-resource language. It contains 17,497 text samples in three classes: Depressive (6,368): texts expressing hopelessness, emotional distress, depressive feelings, or suicidal ideation Non_depressive (5,785): neutral or everyday texts with no depressive indication Positive (5,344): texts expressing appreciation, encouragement, empathy, or positive sentiment File format: one UTF-8 CSV file, BDMentalText_Dataset_V1.csv, with two columns: Text (Bangla text) and Label (class label: Depressive, Non_depressive, or Positive). Characteristics: The texts are informal, written in natural Bangla, and range from short phrases to long narrative posts (2 to over 1,100 words; median 12 words). The vocabulary is large and varied, with about 33,000 unique tokens. The three classes are reasonably balanced, so the data can be used for training and benchmarking without heavy resampling. Depressive texts are much longer on average (about 70 words) than Non_depressive and Positive texts (about 12 words each), so models may learn text length as a shortcut, and users should account for this. A small number of texts (about 1%) appear more than once and should be removed before creating train/test splits. Intended use: Benchmarking machine learning, deep learning, and transformer-based models (e.g., BanglaBERT, XLM-R) for Bangla depression detection, sentiment analysis, and mental-health-related text classification. It can also support research on low-resource NLP, explainable AI, and early detection of distress signals in text. Ethical considerations: The dataset contains sensitive content, including expressions of self-harm and suicidal thoughts. Labels reflect textual expression only and are not clinical diagnoses. The dataset must not be used to diagnose individuals or to make decisions about them, and it is intended for research purposes only. Users should not attempt to identify or contact the authors of any text. If you or someone you know is in distress, please seek help from a qualified professional or local helpline. Access: Because of the sensitive content, the files are restricted. Access can be requested through this Zenodo record, stating your name, affiliation, and intended research use. Citation: If you use this dataset, please cite this Zenodo record.