BanglaVoice(Active, Passive, Middle)

Description

BanglaVoice is a curated sentence-level Bangla dataset annotated for grammatical voice classification into Active, Passive, and Middle constructions. The final annotated dataset contains 4,397 unique Bangla sentences, including 1,459 Active, 1,531 Passive, and 1,407 Middle instances. Each entry includes a cleaned Bangla sentence, its English translation, and the final grammatical voice label. The dataset was compiled from publicly accessible Bangla digital sources, including news portals, magazines, blogs, and literary platforms. A total of 4,957 candidate sentences were initially collected; after preprocessing, 560 sentences were removed because of duplication, incomplete structure, encoding issues, non-linguistic noise, or failure to meet the single-dominant-finite-verb criterion. The final retained dataset contains 4,397 sentences, corresponding to an overall attrition rate of 11.30%. This repository includes the final annotated dataset, per-annotator labels, source-wise distribution and attrition statistics, source-specific collected files, annotation and data collection guidelines, preprocessing and benchmarking scripts, statistical analysis results, and documentation needed to reproduce the dataset construction and baseline classification experiments. Repository structure BanglaVoice/ ├── data/ │ ├── BanglaVoice_final_annotated.xlsx │ ├── BanglaVoice_per_annotator_labels.xlsx │ ├── source_distribution_attrition.xlsx │ ├── Online_News_Corpus.xlsx │ ├── Magazines.xlsx │ ├── Blogs.xlsx │ └── Literary_Sources.xlsx ├── Data collection and annotation guidelines/ │ ├── annotation_guidelines_Readme.md │ └── data_dictionary_Readme.md ├── code/ │ ├── 01_clean_normalize.py │ ├── 02_compute_annotation_agreement.py │ └── 03_benchmark_classification.py ├── results/ │ └── results_statistical_analysis_README.md ├── requirements.txt └── README.md Inter-annotator reliability was measured using Fleiss’ Kappa, with an overall value of κ = 0.955, indicating very high annotation agreement. Sentence-level acceptance was based on agreement by at least two of the three annotators, corresponding to a minimum confidence score of 0.67. The supervised benchmark uses a stratified 80:20 train-test split with random seed 42, character-level TF-IDF features with 3–5 character n-grams, and six classifiers. The best-performing model was LinearSVC, achieving 0.9318 accuracy and 0.9310 F1-score. The dataset is suitable for Bangla grammatical analysis, grammatical voice classification, morpho-syntactic research, semantic role analysis, corpus linguistics, and machine learning applications in Bangla NLP. Value of the Data -Provides a structured sentence-level Bangla corpus annotated for Active, Passive, and Middle voice. -Enables evaluation of Bangla NLP models on voice-sensitive tasks such as semantic role labelling, machine translation, NLG, and question answering.

Steps to reproduce

1. Download all files from the Mendeley Data repository. 2. Install required packages: pip install -r requirements.txt 3. Open the final dataset: data/bangla_voice_unique_rows.xlsx 4. Load the dataset in Python: import pandas as pd df = pd.read_excel("data/bangla_voice_unique_rows.xlsx") 5. To reproduce preprocessing, run: python code/01_clean_normalize.py 6. To recompute annotation agreement, run: python code/02_compute_annotation_agreement.py 7. To reproduce baseline classification, run: python code/03_benchmark_classification.py 8. Use Sentence as the input column and Voice as the target label column.

Institutions

• Daffodil International University

  Dhaka Division, Dhaka

Categories

Licence