• contextual fine-tuning
• deep learning
• context aware machine translation
• machine translation
• state-of-the-art
• neural machine translation
• transformer model
• encoder-decoder architecture
• context adaptation
• human post-editing
• contextual errors
• lexical ambiguities
• real-time adaptation
• translation quality
• human translators
• human-in-the-loop
• BLEU
• COMET

Traditional machine translation industrial systems usually handle sentences independently, neglecting any additional information that may be present beyond the boundaries of the sentence. The focus of this study is to enhance a state-of-the-art Machine Translation system used by human translators by incorporating contextual information only at inference time, while still training the base model using the considerably greater amount of sentence-level data instead of document-level data. In particular, the proposed approach involves a sentence-level fine-tuning at inference time of a general model using contextual information, enabling the system to provide a better in-context translations of sentences. The proposed model has been extensively tested over a wide range of both contextual and non-contextual test sets. The results show that it outperforms the base model in terms of automatic metrics (e.g. BLEU), indicating its superior translation quality. Finally, a human error analysis has been conducted to examine the model's ability to solve contextual errors. The results indicate that the proposed model is effective in resolving lexical ambiguities and producing more coherent translations. Overall, the results demonstrate the efficacy of the proposed approach at a relatively low cost. It can significantly improve the translation quality in real-world situations where the meaning is ambiguous and the context plays a crucial role in determining the correct translation.