Improving Translation Accuracy Through Advanced Language Detection

Effective language identification (LID) is crucial for robust multilingual machine translation (MT), yet current approaches face a critical trade-off between accuracy and efficiency. While static LID methods lack adaptability, joint LID-MT systems introduce computational overhead that limits scalability. This paper presents a novel dynamic LID framework that addresses these challenges by adjusting language representations in real time during translation. Our lightweight module improves MT accuracy while maintaining low latency, particularly benefiting low-resource languages and noisy text scenarios.Experimental results demonstrate significant gains over existing approaches, with +3.5 BLEU improvement for high-resource languages and +6.2 BLEU for low-resource languages, while adding only 8.3% latency overhead a substantial improvement over prior work. The system also achieves 97.5% LID accuracy, outperforming static and joint training baselines. These advances make multilingual MT more practical for real-world applications, where input quality and language variability pose persistent challenges.By integrating dynamic LID with minimal computational cost, this work bridges an important gap between theoretical language representation research and deployable MT systems. The proposed approach opens new directions for adaptive, efficient, and inclusive multilingual translation technologies.