Contextual Code-Mixed Translations with LLM-Based Data Augmentation

Abstract

Code-mixed conversational translation presents unique challenges for neural machine translation (NMT) systems due to its context-dependent nature and linguistic complexity. This paper investigates context-aware translation approaches for Hinglish-to-English code-mixed text, focusing on the comparative performance of the fine-tuned State-Of-The-Art (SOTA) language models, and proposes an approach that incorporates contextual information through a novel preprocessing technique and augmented prompt-based synthetically generated training data. With the help of comprehensive experimentation conducted across distinct configurations using two SOTA Mistral-7B-v0.3 and IndicTrans2 models, the findings demonstrate that Mistral-7Bv0.3 fine-tuned with context-enriched data and synthetic examples achieves state-of-the-art performance with 14.3% improvement over previous approaches. Furthermore, it is observed that in domain-specific patterns in optimal model configuration: conversational data benefits most from contextaware models with synthetic data augmentation, while nonconversational translation performs optimally with syntheticaugmented datasets without contextual enrichment. This research contributes valuable insights into the design of effective translation systems for code-mixed language and establishes new benchmarks for this increasingly important domain of multilingual NLP. © 2025 IEEE.