DIP Project: English-to-Thai Machine Translation system for WIPO Nice Classification Products and Services

📌 Project Overview

This project focuses on English → Thai machine translation (En2Th) with domain-specific emphasis on product and service categories defined by the WIPO Nice Classification system.
The main goal is to evaluate and improve the translation quality for category-specific terminology, ensuring translations are accurate, consistent, and context-aware for intellectual property and business use.

🎯 Objectives

• Benchmark multiple Large Language Models (LLMs) on En2Th and Th2En translation tasks.
• Explore fine-tuning and Retrieval-Augmented Generation (RAG) methods for domain adaptation.
• Provide insights into translation quality across different WIPO Nice product categories.

⚙️ Methodology

1. Data Preparation

   • Source: WIPO Nice classification dataset (product & service categories).
   • Preprocessed into English–Thai parallel pairs.

2. Approaches

   • Baseline Translation: transformer-based (NLLB-200-3.3B) with LoRA fine-tuning
   • Fine-Tuning: Domain-specific fine-tuning on product-category text.
   • RAG (Retrieval-Augmented Generation): Incorporating WIPO classification documents as external knowledge for context-aware translations.

🛠️ Tech Stack

• Fine-tuning: unsloth
• RAG: faiss, langchain
• Inference: vllm
• Evaluation: pythainlp, nltk, jiwer