NLTTA: Nepali Language Transliteration and Translation System

Linguistic Data Science: Solving the Problem

In Nepal, a vast majority of users communicate digitally using informal, Romanized Nepali (e.g., typing "sansad" instead of "parliament"). This heavy code-mixing and transliteration creates a severe barrier for digital inclusion, as most search engines and digital platforms are only optimized for formal English or pure Devanagari text. Users with limited English proficiency often face irrelevant or zero search results. Approaching this as a linguistic data science problem, we recognized that we needed to clean, normalize, and translate this unstructured chat data into a standard target language (English) before it could be indexed and searched efficiently by traditional algorithms.

Research & Dataset

To build a highly specialized translation model, we curated a parallel corpus of Devanagari Nepali, Romanized Nepali, and English. We scraped Wikipedia chunks and combined them with open-source datasets to create a final parallel corpus of 40,000 sentences. The data was rigorously pre-processed, tokenized using SentencePiece, and split into 70% training, 20% validation, and 10% test sets.

System Architecture

Developed using Agile methodologies, our architecture connects state-of-the-art transformer models for NLP and classic probabilistic algorithms for fast document retrieval.

• Translation Engine (mBART Architecture): We fine-tuned mBART (Multilingual Bidirectional and Auto-Regressive Transformer) on our 40,000 sentence parallel corpus. Based on the standard BART architecture, mBART utilizes 12 layers of encoder and 12 layers of decoder. Pre-trained on multiple languages simultaneously using a denoising objective, its powerful cross-lingual transfer capabilities allow it to accurately transliterate and translate informal code-mixed text into standardized English, even in our low-resource setting.
• Search & Retrieval: Once the query is translated, it is passed into a BM25 probabilistic ranking algorithm. BM25 computes relevance scores based on term frequency saturation and document length normalization against a pre-indexed MySQL English news database.
• Deployment: The system operates as a three-tier architecture. A Django REST API acts as the primary orchestrator, interfacing with the ML models and MySQL database, while a Next.js front-end provides the user interface.

Model Training & Evaluation

The mBART model was fine-tuned over 7,000 steps, demonstrating rapid learning convergence. Initially starting at a high loss of 1.1967, the parameters adjusted swiftly to reach a stabilizing training loss of 0.3449.

Handling a morphologically rich, low-resource language required rigorous testing. We evaluated our fine-tuned 630M parameter mBART model against massive state-of-the-art LLMs (like DeepSeek V3 at 671B, ChatGPT at 1.8T, and Llama 4 at 400B) on an unseen test set with highly ambiguous phrases.

• BLEU Score (39.84): Represents highly accurate, publishable quality translation. Astoundingly, our 630M model outperformed the massive LLaMA 2 (70B parameters) specifically in this domain!
• BERT Score (0.9613): Achieved an excellent semantic similarity score, capturing deep contextual meaning robust to synonyms.

While trillion-parameter models naturally scored higher due to their immense scale, our highly focused mBART fine-tune proved vastly superior in terms of computational efficiency, real-time inference speed, and financial scalability—making it the ideal choice for a production-ready search engine in a low-bandwidth environment.

Product Demonstration

The final Notify Nepal web application seamlessly processes complex, code-mixed Romanized Nepali queries and returns accurate English news.