Overview
The HindiChhattisgarhi Machine Translation (MT) System aims to provide reliable, realtime translation between Hindi, the official language of India, and Chhattisgarhi, a regional language spoken by over 18million people in the state of Chhattisgarh. Although Hindi and Chhattisgarhi share the Devanagari script, they differ significantly in vocabulary, morphology, and syntax, making rulebased translation insufficient for fluent output.
This project combines statistical and neural techniques to produce a hybrid MT pipeline that can be used in educational tools, government portals, and mobile apps for citizens who are more comfortable in Chhattisgarhi.
Key Linguistic Challenges
- Lexical divergence: Many everyday words in Chhattisgarhi have no direct Hindi equivalent (e.g., barkha for rain vs. the local barkhaw).
- Morphological richness: Chhattisgarhi uses extensive inflection for case, gender, and honorifics, often encoded as suffixes attached to the verb stem.
- Word order variations: While Hindi follows a canonical SOV order, Chhattisgarhi frequently employs topicalization that moves the object or adverbial phrase to the front of the sentence.
- Codemixing: Speakers frequently insert English or Hindi loanwords, especially for technical terms, complicating vocabulary alignment.
System Architecture
The current implementation consists of three main modules:
1. Preprocessing
- Unicode normalization and script verification.
- Tokenisation using a languagespecific tokenizer that respects compound words and clitics.
- Noise removal (e.g., emojis, URLs) for clean training data.
2. Translation Engine
A hybrid approach is employed:
- Statistical PhraseBased Model (PBSMT): Trained on a parallel corpus of 250k sentence pairs to capture phrase alignments and provide a strong baseline.
- Neural Transformer Model: Built with
fairseq, 6 encoder and 6 decoder layers, 512 hidden size, and 8 attention heads. Finetuned on the same data after PBSMT pretraining. - RuleBased PostProcessing: Handles honorific suffixes and restores gender agreement that the neural model may drop.
3. Postprocessing
- Detokenisation and scriptspecific punctuation handling.
- Confidence scoring using softmax probabilities; lowconfidence sentences are flagged for human review.
- Optional backtranslation for user feedback loops.
Why a hybrid model? PBSMT excels at handling rare words and phrase phrasings that are difficult for neural models with limited data, while the Transformer captures longrange dependencies and fluency. Combining them improves BLEU scores by roughly 34 points over either model alone on the heldout test set.
Data Resources
Building a robust system required gathering and cleaning several datasets:
- Government Gazette Translations: 80k bilingual sentences from official documents.
- Parallel News Articles: 70k sentence pairs scraped from regional news portals that publish in both languages.
- CommunityGenerated Corpus: 100k sentences collected via a crowdsourcing platform where native speakers translated short paragraphs.
- Monolingual Corpora: 3M Hindi and 2M Chhattisgarhi sentences used for languagemodel pretraining and backtranslation.
All data is stored in UTF8 encoded TSV files, with separate files for training, validation, and test splits (80/10/10%).
Evaluation & Results
Evaluation was performed on a blind test set of 5000 sentences using both automatic metrics and human assessment.
Automatic Metrics
| Model | BLEU | TER | ChrF |
|---|---|---|---|
| PBSMT | 27.4 | 58.2 | 0.55 |
| Transformer | 30.1 | 55.3 | 0.58 |
| Hybrid (PBSMT+Transformer) | 33.6 | 51.8 | 0.63 |
Human Evaluation
Twenty bilingual annotators rated 500 randomly selected sentences on a 5point adequacyfluency scale.
- Hybrid system: 4.2 average score (83% satisfactory).
- Transformer alone: 3.7 average score (74% satisfactory).
- PBSMT alone: 3.4 average score (68% satisfactory).
The hybrid model demonstrated notable improvements in handling honorifics and rare lexical items, which were frequent sources of error in the neuralonly version.
Future Directions
- Domain Adaptation: Finetune the model for medical and legal domains using specialized parallel corpora.
- LowResource Augmentation: Employ unsupervised MT techniques and multilingual transfer learning from HindiEnglish models.
- Interactive Learning: Integrate a userfeedback loop in mobile apps to continuously improve the system with realworld corrections.
- Speech Integration: Couple the text MT engine with automatic speech recognition (ASR) and texttospeech (TTS) pipelines to enable spokenlanguage translation.
By expanding resources and incorporating user interaction, the HindiChhattisgarhi MT system can become a vital tool for bridging the digital divide in central India.
