Dataset

This dataset was created by Nuhash Afnan

Contents

BLUGE-TSC: Bangla Sentiment Classification

BLUGE-TSC is a meticulously curated and cleaned Bangla Ternary Sentiment Classification dataset, one of the 7 tasks in BLUGE (Bengali Language UnderstandinG Evaluation), a balanced benchmark for evaluating Bengali natural language understanding. See the full BLUGE collection for all 7 tasks, and the B-CORE pretraining corpus and BnLM model suite released alongside it.

  Dataset Description

This task classifies Bangla text… See the full description on the dataset page: https://huggingface.co/datasets/nahid-hub/BLUGE-bengali-sentiment-classification.

BanglaMUSE is a multimodal Bangla sentiment dataset containing aligned text–audio pairs designed for research in sentiment analysis, speech processing, and multimodal learning for low-resource languages.

The dataset includes 1,000 Bangla sentences, evenly balanced between positive (500) and negative (500) sentiment classes. The sentences represent natural, everyday Bangla language usage and were manually curated and validated to ensure clear sentiment polarity.

Each sentence is recorded by four native Bangla speakers (two female and two male), resulting in 4,000 speech recordings in total. All speakers recorded the same set of sentences, enabling controlled analysis of speaker variability while preserving identical textual content. Audio samples are provided in MP3 format, recorded in controlled indoor environments, and manually verified for quality and alignment.

The dataset is distributed with a unified metadata.csv file that links sentence identifiers, sentiment labels, speaker information, and relative audio paths. BanglaMUSE supports tasks such as multimodal sentiment classification, sentiment-aware speech recognition, audio–text alignment, and speaker-independent modeling.

This is a data set of Sentiment Analysis On Bangla News Comments where every data was annotated by three different individuals to get three different perspectives and based on the majorities decisions the final tag was chosen. This data set contains 13802 data in total.

This dataset comprises 3,290 Bengali political comments sourced from social media platforms, news comment sections, and online political discussions, specifically curated for sentiment analysis research in Bengali NLP. The corpus provides a comprehensive resource for training and evaluating sentiment classification models within the political domain. The dataset features 3,290 instances distributed across five sentiment classes with excellent balance (variance <8%): Very Negative (675, 20.5%), Negative (663, 20.2%), Neutral (626, 19.0%), Very Positive (664, 20.2%), and Positive (662, 20.1%). Stored in Excel format with two columns containing Bengali political comments (Unicode text) and corresponding sentiment labels, the dataset maintains high quality with no missing values and verified annotations. Comment lengths average 83 characters, ranging from 11 to 398 characters. The collection encompasses diverse political discourse including government policies and governance, electoral processes and democracy, political parties and leadership dynamics, social and economic issues, current affairs and political events, along with public opinion and citizen responses to political developments. This dataset serves multiple research purposes, including Bengali sentiment analysis model development and benchmarking, political discourse analysis and opinion mining, natural language processing research for low-resource languages, cross-lingual sentiment analysis studies, social media analytics for Bengali content, multi-class text classification research, and comparative political sentiment studies across different linguistic and cultural contexts.

SentiFive is a multi-class Bengali sentiment analysis dataset consisting of 31,411 YouTube comments, manually annotated into five sentiment categories: Strongly Negative, Weakly Negative, Neutral, Weakly Positive, and Strongly Positive. The dataset is designed to support research in fine-grained sentiment classification and low-resource language processing.

Unlike previous Bengali sentiment datasets that focus on binary or ternary sentiment, SentiFive enables more nuanced modeling of user opinions expressed in informal social media contexts. Data were collected from a diverse set of YouTube videos, covering topics such as news, entertainment, and politics.

M. A. Rahman, A. Mahbub, B. N. Paul, P. Bhattacharjee and M. A. -U. -Z. Ashik, "SentiFive: A Multi-Class Bengali Dataset for Sentiment Analysis," 2025 IEEE 7th International Conference on Sustainable Technologies For Industry 5.0 (STI), Dhaka, Bangladesh, 2025, pp. 1-6, doi: 10.1109/STI69347.2025.11367591. keywords: {Deep learning;Sentiment analysis;Video on demand;Social networking (online);Bidirectional long short term memory;Web sites;Reliability;Fifth Industrial Revolution;Standards;Software development management;SentiFive;5-Class Sentiment;Baseline Evaluation;LSTM;BiLSTM;Bangla Natural Language Processing (BNLP);Sentiment Analysis (SA)},

Dataset

This dataset was created by Rhs Liza

Released under CC0: Public Domain

Contents

The BanglaMUSE-VID dataset is publicly available through the Science Data Bank (SciDB) and provides a synchronized multimodal benchmark for Bangla sentiment analysis. The repository contains 1,000 manually curated Bangla sentences with balanced sentiment labels (500 positive and 500 negative) and 5,000 corresponding face-visible MP4 video recordings captured by five native Bangla speakers using different smartphone devices. The dataset comprises both textual and video modalities, enabling research in multimodal sentiment analysis, visual speech recognition, cross-modal representation learning, and video-grounded language understanding for low-resource Bangla NLP applications. The repository is openly accessible and designed for future extensibility through the inclusion of additional speakers, sentiment categories, and domain-specific content.

The total amount of collected comments is 44001. The dataset aims to differentiate whether a comment is a bully expression or not with the help of Natural Language Processing and to what extent it is improper if it is an inappropriate comment. The comments are labeled with different categories of harassment with the help of experts and consensus.

ANUBHUTI, a comprehensive dataset consisting of 2,500 sentences manually translated from standard Bangla into four major regional dialects—Mymensingh, Noakhali, Sylhet, and Chittagong. The dataset predominantly features political and religious content, reflecting the contemporary socio-political landscape of Bangladesh, alongside neutral texts to maintain balance. Each sentence is annotated using a dual annotation scheme: (i) multiclass thematic labeling categorizes sentences as Political, Religious, or Neutral, and (ii) multilabel emotion annotation assigns one or more emotions from Anger, Contempt, Disgust, Enjoyment, Fear, Sadness, and Surprise.

Contains training and test sets as well as the manually created Twitter-specific lexicons for the following paper,

Performing Sentiment Analysis in Bangla Microblog Posts

The resources are made available to encourage more research in Bangla sentiment classification and other NLP tasks.

Please cite the paper if you use the dataset or lexicon

This is a dataset for the sentiment analysis task. It contains 100k code mixed data. The languages are Bangla-English-Hindi.

Dataset Generation:

Initially, we select the Amazon Review Dataset as our base data, referenced from Ni et al. (2019)**[1]**. We randomly extract 100,000 instances from this dataset. The original labels in this dataset are ratings, scaled from 1 to 5. For our specific task, we categorize them into Positive (rating > 3), Neutral (rating = 3), and Negative (rating < 3), ensuring a balanced number of instances for each label. To generate the synthetic Code-mixed dataset, we apply two distinct methodologies: the Random Code-mixing Algorithm by Krishnan et al. (2021)**[2]** and r-CM by Santy et al. (2021)**[3]**.

Class Distribution:

For train.csv:

For dev.csv:

For test.csv:

Cite our Paper:

If you utilize this dataset, kindly cite our paper.

@article{raihan2023mixed, title={Mixed-Distil-BERT: Code-mixed Language Modeling for Bangla, English, and Hindi}, author={Raihan, Md Nishat and Goswami, Dhiman and Mahmud, Antara}, journal={arXiv preprint arXiv:2309.10272}, year={2023} }

References

[1]: Ni, J., Li, J., & McAuley, J. (2019). Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP) (pp. 188-197).

[2]: Krishnan, J., Anastasopoulos, A., Purohit, H., & Rangwala, H. (2021). Multilingual code-switching for zero-shot cross-lingual intent prediction and slot filling. arXiv preprint arXiv:2103.07792.

[3]: Santy, S., Srinivasan, A., & Choudhury, M. (2021). BERTologiCoMix: How does code-mixing interact with multilingual BERT? In Proceedings of the Second Workshop on Domain Adaptation for NLP (pp. 111-121).

ANUBHUTI, a comprehensive dataset consisting of 2,000 sentences manually translated from standard Bangla into four major regional dialects—Mymensingh, Noakhali, Sylhet, and Chittagong. The dataset predominantly features political and religious content, reflecting the contemporary socio-political landscape of Bangladesh, alongside neutral texts to maintain balance. Each sentence is annotated using a dual annotation scheme: (i) multiclass thematic labeling categorizes sentences as Political, Religious, or Neutral, and (ii) multilabel emotion annotation assigns one or more emotions from Anger, Contempt, Disgust, Enjoyment, Fear, Sadness, and Surprise.

Dataset

This dataset was created by M Arman

Contents

Motamot: Bengali Political Sentiment Analysis Dataset

📖 Overview

Motamot is a Bengali political sentiment analysis dataset containing 7,058 labeled data points. Each entry is annotated with Positive or Negative sentiment, specifically tailored for analyzing political discourse in the Bengali language.

This dataset supports Natural Language Processing (NLP) research, with applications in sentiment classification, political opinion mining, and benchmarking pre-trained and large language models (LLMs) for low-resource languages.

📊 Dataset Statistics

📜 Citation

If you use this dataset, please cite the following paper:

@INPROCEEDINGS{10752197,
 author={Johora Faria, Fatema Tuj and Moin, Mukaffi Bin and Mumu, Rabeya Islam and Alam Abir, Md Mahabubul and Alfy, Abrar Nawar and Alam, Mohammad Shafiul},
 booktitle={2024 IEEE Region 10 Symposium (TENSYMP)}, 
 title={Motamot: A Dataset for Revealing the Supremacy of Large Language Models Over Transformer Models in Bengali Political Sentiment Analysis}, 
 year={2024},
 pages={1-8},
 keywords={Sentiment analysis;Analytical models;Accuracy;Voting;Large language models;Transformers;Market research;Few shot learning;Portals;IEEE Regions;Political Sentiment Analysis;Pre-trained Language Models;Large Language Models;Gemini 1.5 Pro;GPT 3.5 Turbo;Zero-shot Learning;Fewshot Learning;Low-resource Language},
 doi={10.1109/TENSYMP61132.2024.10752197}
}

📂 Dataset Structure

Motamot/
│
├── train.csv    # Training set (5,647 instances)
├── test.csv     # Test set (706 instances)
├── validation.csv  # Validation set (705 instances)

Each file contains:

• text → Bengali political statement
• label → Sentiment category (Positive or Negative)

🧪 Baseline Results

🔹 Comparative Analysis of Pre-trained Language Models

🔹 Comparative Analysis of Large Language Models (Few-shot & Zero-shot)

📬 Contact Information

For questions, collaborations, or inquiries:

• Mukaffi Bin Moin — 📧 mukaffi28@gmail.com
• Fatema Tuj Johora Faria — 📧 fatema.faria142@gmail.com
• Rabeya Islam Mumu — 📧 rabeya.islammomo@gmail.com
• Md Mahabubul Alam Abir — 📧 mahbubabir09@gmail.com
• Abrar Nawar Alfy — 📧 abraralfy49@gmail.com

The dataset consists of 34,812 Bengali posts and comments sourced from Facebook, Twitter, and Instagram, Bengali news portals and literature. Techniques employed in data acquisition included data scraping from social media accounts through API and scraping only text data from websites. Microblogs consist of posts and comments from platforms like Facebook, Twitter, and Instagram, which allow for the capture of informal and emotionally rich text. Newspaper and magazine articles provide formal, sentiment-related information through opinions. Online literature, including Bengali novels, poems, and blogs, incorporates semantic relationships and linguistic nuances. Text data is collected from public sources through automated scripts. We used selenium scripts, created using the Python programming language. We used APIs to obtain structured social media data. Additionally, we complied with the requirements of privacy, data collection, and ethics.It contains 5 Emotion and 5 Sentiment class. For emotion "Creepy" being the most frequent emotion with 12,000 entries, followed by "Unbiased" with 8,500 entries, "Joyful" with 7,500 entries, "Bullying" with 4,000 entries, and "Surprise" with 2,500 entries. On the other hand, for sentiment "Negative" being the most frequent with 8,000 entries, followed by "Neutral" with 7,000 entries, "Strongly Negative" with 6,800 entries, "Positive" with 5,500 entries, and "Strongly Positive" with 4,500 entries in that order.

This BanglaBlend dataset is a comprehensive collection of Bangla (Bengali) sentences meticulously categorized based on two specific forms: Saint(Sadhu) and Common(Cholito). This dataset is comprised of a total 7350 annotated Bangla sentences as well as it is preprocessed dataset where several data preprocessing techniques have been applied. This dataset is designed to facilitate research and development in natural language processing (NLP) and computational linguistics, particularly for Bangla, a widely spoken language in Bangladesh and parts of India. Specially, this dataset can be leveraged for several natural language processing task such as text summarization, text classification, sentiment analysis, automatic language translation.

In recent years, the surge in reviews and comments on newspapers and social media has made sentiment analysis a focal point of interest for researchers. Sentiment analysis is also gaining popularity in the Bengali language. However, Aspect-Based Sentiment Analysis is considered a difficult task in the Bengali language due to the shortage of perfectly labeled datasets and the complex variations in the Bengali language. This study used two open-source benchmark datasets of the Bengali language, Cricket, and Restaurant, for our Aspect-Based Sentiment Analysis task. The original work was based on the Random Forest, Support Vector Machine, K-Nearest Neighbors, and Convolutional Neural Network models. In this work, we used the Bidirectional Encoder Representations from Transformers, the Robustly Optimized BERT Approach, and our proposed hybrid transformative Random Forest and Bidirectional Encoder Representations from Transformers (tRF-BERT) models to compare the results with the existing work. After comparing the results, we can clearly see that all the models used in our work achieved better results than any of the previous works on the same dataset. Amongst them, our proposed transformative Random Forest and Bidirectional Encoder Representations from Transformers achieved the highest F1 score and accuracy. The accuracy and F1 score of aspect detection for the Cricket dataset were 0.89 and 0.85, respectively, and for the Restaurant dataset were 0.92 and 0.89 respectively.

Dataset

This dataset was created by Mahfuz Ahmed Masum

Released under CC0: Public Domain

Contents

Welcome to our Bengali Financial News Sentiment Analysis dataset! This collection comprises 7,695 financial news articles extracted, covering the period from March 3, 2014, to December 29, 2021. Utilizing the powerful web scraping tool "Beautiful Soup 4.4.0" in Python.

This dataset was a crucial part of our research published in the journal paper titled "Stock Market Prediction of Bangladesh Using Multivariate Long Short-Term Memory with Sentiment Identification." The paper can be accessed and cited at http://doi.org/10.11591/ijece.v13i5.pp5696-5706.

We are excited to share this unique dataset, which we hope will empower researchers, analysts, and enthusiasts to explore and understand the dynamics of the Bengali financial market through sentiment analysis. Join us on this journey of uncovering the hidden emotions driving market trends and decisions in Bangladesh. Happy analyzing!