AI-102 Question 236
Select 2You are tasked with creating a custom translation model using Azure Translator to translate technical documents from English to German. After uploading your training and tuning data, you notice that the model's performance is not meeting your accuracy requirements. What steps should you take to improve and publish the custom model?
- A
Add more high-quality parallel sentence pairs to your training dataset.
- B
Use the Translator's pre-trained generic model as a fallback during the custom model's training.
- C
Refine your tuning dataset to ensure it represents the specific domain or context of your translation.
- D
Directly publish the model without further changes to observe its performance in real-world usage.
- E
Retrain the model with a larger and more diverse dataset, even if it is not domain-specific.
Show answer and explanation
Correct answers: A, C
Explanation
To improve the performance of a custom translation model, it is essential to provide high-quality and domain-specific training data, as well as a well-representative tuning dataset. Adding more high-quality parallel sentence pairs and refining the tuning dataset are the most effective strategies to enhance the model's accuracy and relevance for the technical domain before publishing.
- A. Correct.
Adding more high-quality parallel sentence pairs improves the training dataset, helping the model learn better correlations between the source and target languages.
- B. Incorrect.
The Translator's pre-trained generic model cannot be explicitly used as a fallback during custom model training. Custom models are trained independently from pre-trained models.
- C. Correct.
Refining the tuning dataset ensures that the model's evaluation aligns with the specific domain or context, which is crucial for achieving better translation accuracy.
- D. Incorrect.
Publishing the model without further changes would not improve its current accuracy. It is better to address the issues with training and tuning first.
- E. Incorrect.
Retraining with a larger but non-domain-specific dataset may dilute the domain-specific accuracy of the model, which is critical for translating technical documents.