BHT25: Bengali-Hindi-Telugu Parallel Corpus for Literary Machine Translation

Overview

BHT25 is a high-quality trilingual parallel corpus comprising 27,149 sentence triplets across Bengali (BN), Hindi (HI), and Telugu (TE) languages. This dataset addresses a critical gap in resources for cross-family Indian language machine translation, particularly for literary and culturally rich content spanning Indo-Aryan and Dravidian language families.

Key Features:

• 🌐 Three Indian Languages: Bengali (Indo-Aryan), Hindi (Indo-Aryan), Telugu (Dravidian)
• 📚 Literary Domain: Sourced from renowned authors including Rabindranath Tagore and Sarat Chandra Chattopadhyay
• ✨ Archaic Varieties: Includes traditional Bengali Sadhu Bhasha for diachronic NLP research
• ✅ Human-Verified: 75.8% semantic alignment accuracy validated by expert linguists
• 🔍 Unique Identifiers: Each triplet has a unique ID (BHT25_XXXXX) for reproducibility
• 📖 High Quality: Composite fluency score of 4.08/5.0 from native speaker ratings
• 🔓 Open Access: Released under CC BY 4.0 license

Dataset Description

Languages

Language	Family	Script	Speakers	ISO 639-1
Bengali	Indo-Aryan	Bengali (Bangla)	~265M	bn
Hindi	Indo-Aryan	Devanagari	~600M	hi
Telugu	Dravidian	Telugu	~95M	te

Dataset Statistics

Metric	Bengali (bn)	Hindi (hi)	Telugu (te)
Total Sentences	27,149	27,149	27,149
Total Tokens	~420,000	~386,000	~445,000
Avg. Tokens/Sentence	15.5 ± 8.3	14.3 ± 7.6	16.4 ± 9.1
Avg. Characters/Sentence	87.3 ± 48.2	82.1 ± 45.7	95.7 ± 52.6
Vocabulary Size	~47,856	~43,214	~52,143
Min Sentence Length	3 tokens	3 tokens	3 tokens
Max Sentence Length	147 tokens	132 tokens	156 tokens
Median Sentence Length	14 tokens	13 tokens	15 tokens

Note: Telugu exhibits slightly longer average sentence lengths due to its agglutinative morphology.

Content Characteristics

The corpus encompasses diverse literary genres to ensure broad applicability:

• Narrative Fiction (45.2%): Short stories and novel excerpts
• Poetry and Verse (18.7%): Traditional and modern poetry
• Folk Literature (15.6%): Folk tales and oral traditions
• Contemporary Prose (12.3%): Modern literary essays and articles
• Classical Literature (8.2%): Traditional Sadhu Bhasha texts

Quality Metrics

Quality Measure	Score	Method
Alignment Accuracy	75.8%	Human validation (500-sample random subset)
Semantic Consistency (bn-te)	0.873	Cross-lingual Word Embeddings (CLWE) similarity
Semantic Consistency (bn-hi)	0.81	CLWE similarity
Semantic Consistency (hi-te)	0.71	CLWE similarity
Translation Fluency	4.08/5.0	Expert annotation (composite score, n=3 per language)
Inter-Annotator Agreement	κ = 0.89	Fleiss' kappa coefficient

Lower semantic consistency between Hindi and Telugu reflects greater typological distance between Indo-Aryan and Dravidian families.

Dataset Structure

Data Format

The dataset is provided in Apache Parquet format with the following schema:

{
    'id': string,        # Unique identifier (BHT25_00001 to BHT25_27149)
    'bn': string,        # Bengali sentence (UTF-8 encoded)
    'hi': string,        # Hindi sentence (UTF-8 encoded)
    'te': string         # Telugu sentence (UTF-8 encoded)
}

Example Triplets

{
    'id': 'BHT25_00001',
    'bn': 'হুগলি জেলার সপ্তগ্রামে দুই ভাই নীলাম্বর ও পীতাম্বর চক্রবর্তী বাস করিত',
    'hi': 'हुगली जिले का सप्तग्राम-उसमें दो भाई नीलाम्बर व पीताम्बर रहते थे',
    'te': 'హుగ్లీ జిల్లాలోని సప్తగ్రామ్-దీనికి ఇద్దరు సోదరులు నీలాంబర్ మరియు పితాంబర్ అక్కడ నివసించేవారు.'
}

{
    'id': 'BHT25_00015',
    'bn': 'আজ সকালে নীলাম্বর চন্ডীমণ্ডপের একধারে বসিয়া তামাক খাইতেছিল',
    'hi': 'आज सवेरे नीलाम्बर चण्डी-मण्डप में बैठा हुक्का पी रहा था',
    'te': 'ఈ ఉదయం నీలాంబర్ చండీ-మండపంలో కూర్చుని హుక్కా తాగుతున్నాడు.'
}

Data Splits

The dataset is provided as a single unified corpus without pre-defined train/development/test splits. This design choice maximizes research flexibility, allowing users to:

• Create custom split ratios (80-10-10, 70-15-15, 90-5-5, etc.)
• Implement k-fold cross-validation
• Combine with other datasets
• Use deterministic splitting via unique IDs

Suggested Split Strategy (for standardization):

• Train: 80% (21,719 triplets)
• Development: 10% (2,715 triplets)
• Test: 10% (2,715 triplets)

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("sudeshna84/BHT25")

# Access data
print(f"Total samples: {len(dataset['train'])}")
print(f"First example:\n{dataset['train'][0]}")

# Example output:
# Total samples: 27149
# First example:
# {
#   'id': 'BHT25_00001',
#   'bn': 'হুগলি জেলার সপ্তগ্রামে দুই ভাই নীলাম্বর ও পীতাম্বর চক্রবর্তী বাস করিত',
#   'hi': 'हुगली जिले का सप्तग्राम-उसमें दो भाई नीलाम्बर व पीताम्बर रहते थे',
#   'te': 'హుగ్లీ జిల్లాలోని సప్తగ్రామ్-దీనికి ఇద్దరు సోదరులు నీలాంబర్ మరియు పితాంబర్ అక్కడ నివసించేవారు.'
# }

Creating Train/Dev/Test Splits

from datasets import load_dataset

# Load dataset
dataset = load_dataset("sudeshna84/BHT25", split="train")

# Create 80-10-10 split (reproducible with seed)
train_test = dataset.train_test_split(test_size=0.2, seed=42)
train_dataset = train_test['train']  # 21,719 triplets
temp_dataset = train_test['test']     # 5,430 triplets

# Further split test into dev and test
dev_test = temp_dataset.train_test_split(test_size=0.5, seed=42)
dev_dataset = dev_test['train']   # 2,715 triplets
test_dataset = dev_test['test']   # 2,715 triplets

print(f"Train: {len(train_dataset)}, Dev: {len(dev_dataset)}, Test: {len(test_dataset)}")
# Output: Train: 21719, Dev: 2715, Test: 2715

Accessing Specific Language Pairs

# Extract Bengali-Hindi pairs
bn_hi_pairs = [(item['bn'], item['hi']) for item in dataset['train']]
print(f"Bengali-Hindi pairs: {len(bn_hi_pairs)}")

# Extract Bengali-Telugu pairs
bn_te_pairs = [(item['bn'], item['te']) for item in dataset['train']]
print(f"Bengali-Telugu pairs: {len(bn_te_pairs)}")

# Extract Hindi-Telugu pairs
hi_te_pairs = [(item['hi'], item['te']) for item in dataset['train']]
print(f"Hindi-Telugu pairs: {len(hi_te_pairs)}")

Integration with Translation Models

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from datasets import load_dataset

# Load dataset
dataset = load_dataset("sudeshna84/BHT25", split="train")

# Example: Load IndicTrans2 model for Bengali→Hindi translation
tokenizer = AutoTokenizer.from_pretrained("ai4bharat/indictrans2-en-indic-dist-200M")
model = AutoModelForSeq2SeqLM.from_pretrained("ai4bharat/indictrans2-en-indic-dist-200M")

# Prepare data for fine-tuning
def preprocess_function(examples):
    # IndicTrans2 requires language tags
    inputs = [f"ben_Beng: {text}" for text in examples['bn']]
    targets = [f"hin_Deva: {text}" for text in examples['hi']]
    
    model_inputs = tokenizer(
        inputs, 
        max_length=128, 
        truncation=True, 
        padding='max_length'
    )
    
    labels = tokenizer(
        targets, 
        max_length=128, 
        truncation=True, 
        padding='max_length'
    )
    
    model_inputs["labels"] = labels["input_ids"]
    return model_inputs

# Tokenize dataset
tokenized_dataset = dataset.map(
    preprocess_function, 
    batched=True,
    remove_columns=dataset.column_names
)

# Now ready for training with HuggingFace Trainer!

Data Analysis Example

import pandas as pd
import matplotlib.pyplot as plt
from datasets import load_dataset

# Load dataset
dataset = load_dataset("sudeshna84/BHT25", split="train")

# Convert to pandas DataFrame
df = pd.DataFrame(dataset)

# Analyze sentence lengths (in tokens)
df['bn_length'] = df['bn'].str.split().str.len()
df['hi_length'] = df['hi'].str.split().str.len()
df['te_length'] = df['te'].str.split().str.len()

# Print statistics
print("Sentence Length Statistics (tokens):")
print(df[['bn_length', 'hi_length', 'te_length']].describe())

# Example output:
#        bn_length  hi_length  te_length
# count  27149.00   27149.00   27149.00
# mean      15.48      14.27      16.39
# std        8.31       7.58       9.12
# min        3.00       3.00       3.00
# 25%       10.00       9.00      10.00
# 50%       14.00      13.00      15.00
# 75%       19.00      18.00      21.00
# max      147.00     132.00     156.00

# Visualize length distributions
fig, axes = plt.subplots(1, 3, figsize=(15, 4))
df['bn_length'].hist(bins=50, ax=axes[0], color='skyblue', edgecolor='black')
axes[0].set_title('Bengali Sentence Length Distribution')
axes[0].set_xlabel('Tokens')
axes[0].set_ylabel('Frequency')

df['hi_length'].hist(bins=50, ax=axes[1], color='lightcoral', edgecolor='black')
axes[1].set_title('Hindi Sentence Length Distribution')
axes[1].set_xlabel('Tokens')

df['te_length'].hist(bins=50, ax=axes[2], color='lightgreen', edgecolor='black')
axes[2].set_title('Telugu Sentence Length Distribution')
axes[2].set_xlabel('Tokens')

plt.tight_layout()
plt.savefig('sentence_length_distribution.png', dpi=300)
plt.show()

Applications

This corpus supports a wide range of NLP research:

1. Machine Translation

• Neural MT training and evaluation: Fine-tune mBART, mT5, IndicTrans2
• Cross-family translation: Study challenges in Indo-Aryan ↔ Dravidian transfer
• Low-resource language pairs: Bootstrap Hindi-Telugu models via Bengali pivot
• Domain adaptation: Literary translation quality assessment

2. Cross-Lingual Analysis

• Cross-lingual word embeddings: Evaluation on semantic similarity tasks
• Syntactic divergence analysis: Compare word order and morphological strategies
• Translation quality estimation: Automatic metric development for Indian languages

3. Multilingual NLP

• Multilingual language model fine-tuning: BERT, XLM-R, mBERT
• Zero-shot translation: Transfer learning across language families
• Multilingual sentiment/emotion analysis: Literary emotion preservation

4. Linguistic Research

• Comparative morphology: Agglutinative (Telugu) vs. inflectional (Bengali/Hindi)
• Literary translation studies: Style and cultural adaptation analysis
• Diachronic NLP: Archaic Bengali (Sadhu Bhasha) processing
• Typological studies: Word order, case marking, verb morphology

5. Quality Estimation

• Alignment algorithm evaluation: Benchmark Gale-Church, CLWE-based methods
• Translation quality metrics: BLEU, chrF, METEOR for Indic languages
• Human evaluation correlation: Automatic metrics vs. expert ratings

Methodology

Data Collection Pipeline

┌─────────────────────────┐
│  Literary Text Sources  │
│  (Tagore, Sarat Chandra,│
│   Folk Literature, etc.)│
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  Digitization & OCR     │
│  (Google Cloud Vision,  │
│   Tesseract 4.1)        │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  Text Preprocessing     │
│  (Unicode normalization,│
│   script standardization)│
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  Sentence Segmentation  │
│  (Language-specific     │
│   punctuation rules)    │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  Trilingual Alignment   │
│  - Gale-Church (length) │
│  - CLWE similarity      │
│  - Hybrid scoring       │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  Quality Validation     │
│  - Automatic filtering  │
│  - Expert review (n=9)  │
│  - Fluency rating       │
│  - IAA: κ=0.89          │
└───────────┬─────────────┘
            │
            ▼
┌─────────────────────────┐
│  Final Corpus (27,149)  │
│  Format: Parquet        │
│  Unique IDs assigned    │
└─────────────────────────┘

Alignment Algorithm

The corpus employs a hybrid alignment approach:

1. Gale-Church Length-Based Alignment:

   • Character-count ratios optimized for Indian languages
   • Adjusted for Telugu agglutinative morphology

2. CLWE Semantic Refinement:

   • FastText multilingual embeddings (300-dim)
   • Cosine similarity threshold: ≥0.7

3. Manual Validation:

   • 500-sample random verification
   • Inter-annotator agreement: κ=0.89

Quality Assurance

Three-stage validation:

• Stage 1: Automatic filtering (length, encoding, language detection)
• Stage 2: Expert review by 9 native speakers (3 per language)
• Stage 3: Iterative refinement based on consensus

Fluency Rating Scale:

• 5: Perfectly natural and fluent
• 4: Minor awkwardness, generally fluent
• 3: Understandable but noticeable issues
• 2: Significant fluency problems
• 1: Incomprehensible or severely malformed

Mean corpus fluency: 4.08 ± 0.67

Citation

If you use this dataset in your research, please cite:

@article{sani2024bht25,
  title={BHT25: A Bengali-Hindi-Telugu Parallel Corpus for Enhanced Literary Machine Translation},
  author={Sani, Sudeshna and Gangashetty, Suryakanth V and Samudravijaya, K and Nandi, Anik and Priya, Aruna and Kumar, Vineeth and Dubey, Akhilesh Kumar},
  journal={Data in Brief},
  year={2024},
  volume={XX},
  pages={XXXXX},
  publisher={Elsevier},
  doi={10.1016/j.dib.2024.XXXXX}
}

Related Research (ESA-NMT Model):

@ARTICLE{11333267,
  author={Sani, Sudeshna and Gangashetty, Suryakanth V. and Samudravijaya, K. and Dubey, Akhilesh Kumar},
  journal={IEEE Access}, 
  title={Emotion-Semantic-Aware Neural Machine Translation Between Indo-Aryan and Dravidian Languages via Transfer Learning}, 
  year={2026},
  volume={14},
  number={},
  pages={4953-4969},
  keywords={Translation;Semantics;Cultural differences;Transfer learning;Multilingual;Transformers;Electronic mail;Standards;Neural machine translation;Accuracy;Bengali-Hindi-Telugu;emotion recognition;literary translation;neural machine translation;semantic consistency;transfer learning},
  doi={10.1109/ACCESS.2026.3651419}}

License

This dataset is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0).

You are free to:

• ✅ Share: Copy and redistribute the material in any medium or format
• ✅ Adapt: Remix, transform, and build upon the material for any purpose, even commercially

Under the following terms:

• Attribution: You must give appropriate credit, provide a link to the license, and indicate if changes were made. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use.

Ethical Considerations

Copyright Compliance

• All source texts are either public domain (published >95 years ago) or used with explicit permission
• Author attributions maintained in metadata
• No unauthorized copyrighted material included

Annotator Welfare

• Fair compensation (₹500/hour) for all human annotators
• Work limited to 4 hours/day to prevent fatigue
• Option to decline sensitive content

Content Screening

• No explicit sexual content
• No graphic violence or hate speech
• No personally identifiable information
• 0.3% of extracted sentences excluded on ethical grounds (78 triplets)

Data Privacy

• All names in corpus are fictional literary characters or historical public figures
• No user-generated content or social media data
• Full compliance with data protection regulations

Known Limitations

Alignment Quality

• 5.8% of triplets have suboptimal alignment (validation score = 0)
• Primarily occurs in complex literary passages with metaphorical language
• Quality scores provided in metadata for user-aware filtering

Genre Imbalance

• Corpus skews toward narrative fiction (45.2%)
• Poetry and technical writing underrepresented
• Users should account for domain bias in downstream tasks

Archaic Language

• Bengali Sadhu Bhasha (8.2%) differs from contemporary Bengali
• May pose challenges for modern MT systems
• Valuable for diachronic NLP research
• Can be filtered using genre metadata if needed

Residual OCR Errors

• Estimated error rate: <0.5% per language
• Most common in rare characters and conjuncts
• Users encouraged to report errors via GitHub issues

Contributing

We welcome contributions to improve the dataset quality:

Reporting Issues

• GitHub Issues: github.com/sudeshna84/BHT25
• Email: sudeshna.sani@klef.edu.in

Corrections and Improvements

• Submit pull requests with specific triplet corrections
• Provide evidence (source text citations) for proposed changes
• All contributions will be reviewed and credited

Authors and Affiliations

Sudeshna Sani¹ (Corresponding Author)
Suryakanth V Gangashetty¹
Samudravijaya K¹
Anik Nandi²
Aruna Priya³
Vineeth Kumar³
Akhilesh Kumar Dubey¹

¹ Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation (KLEF), Guntur, Andhra Pradesh, India
² School of Business, Woxsen University, Hyderabad, Telangana, India
³ School of Technology, Woxsen University, Hyderabad, Telangana, India

Contact

Corresponding Author: Sudeshna Sani
📧 Email: sudeshna.sani@klef.edu.in
🔗 GitHub: github.com/sudeshna84
🏢 Affiliation: KLEF, Guntur, India

For dataset-specific inquiries:

• Open an issue in this repository
• Email the corresponding author
• Use HuggingFace Discussions tab

Acknowledgments

We gratefully acknowledge:

• Native speakers who participated in manual validation (9 expert annotators)
• Literary estates for permission to use copyrighted source materials
• India's National Translation Mission for inspiring this work
• AI4Bharat initiative for Indic NLP tools and resources
• KLEF and Woxsen University for institutional support

Changelog

Version 1.0 (December 2024) - Initial Release

• 27,149 sentence triplets
• Three languages: Bengali, Hindi, Telugu
• Quality metrics: 75.8% alignment accuracy, 4.08/5.0 fluency
• Parquet format with unique IDs
• Complete metadata and documentation

Roadmap

Future enhancements planned:

• v1.1 (Q1 2025): Add sentence-level emotion annotations
• v2.0 (Q3 2025): Expand to 50,000 triplets
• v2.1 (Q4 2025): Include parallel audio for multilingual speech research
• v3.0 (2026): Extend to Tamil and Kannada (pan-South-Indian coverage)

---

Dataset Version: 1.0
Last Updated: December 19, 2024
DOI: [Will be assigned upon Data in Brief publication]
HuggingFace Downloads:

---

For comprehensive methodology details, statistical analyses, and validation procedures, please refer to our Data in Brief paper (citation above).