AI-900 Practice Question: Describe features of Natural Language Processing workloads on Azure

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Automatically translating training data from English to other languages before fine-tuning

  Why it's wrong here

  Automatically translating training data is a data augmentation technique that superficially expands the dataset, but it does not exploit the model's internal cross-lingual representation. Cross-lingual models like XLM-R are pre-trained on massive multilingual corpora and learn to map similar concepts across languages into aligned vector spaces, so fine-tuning on English alone naturally generalizes. In contrast, translated data adds noise, can introduce literalness, and still treats each language as a separate task rather than sharing parameters. Thus, translation before fine-tuning is an alternative approach, not the core mechanism of cross-lingual transfer.

• ✓

  Using shared multilingual representations so knowledge learned in one language transfers to others

  Why this is correct

  Cross-lingual models, such as multilingual BERT or XLM-R, are trained on many languages simultaneously with a shared dictionary and transformer encoder. When fine-tuned on English task data, the gradient updates adjust shared parameters that also affect representations for other languages, allowing the model to apply the learned task knowledge to those languages. This shared multilingual representation space is precisely why knowledge transfers across languages without needing parallel data or translation. Therefore, this option correctly describes the fundamental mechanism underlying cross-lingual transfer.

• ✗

  Using the same model for both NLP and computer vision tasks

  Why it's wrong here

  Using a single model for both NLP and computer vision requires a multi-modal architecture, such as CLIP or Flamingo, that aligns image and text embeddings, which is fundamentally different from cross-lingual transfer. Cross-lingual transfer operates exclusively within language, allowing knowledge learned on English, for example, to be applied to French or Hindi without retraining. A multi-modal model does not address multilingual sharing; it addresses how different modalities map to a shared semantic space. Therefore, this option describes a different capability and fails the test of what cross-lingual transfer means.

• ✗

  Transferring a model trained in Azure to run on another cloud provider

  Why it's wrong here

  Transferring a trained model between cloud providers is an infrastructure and deployment concern—handling serialization, containerization, and API compatibility—rather than a machine learning technique. Cross-lingual transfer is about the model's ability to generalize across languages through shared multilingual parameters, which has nothing to do with where the model runs. Cloud portability doesn't alter the model's weights or vocabulary, so it cannot enable knowledge learned in one language to benefit another. This option confuses the runtime environment with the model's linguistic capabilities.

Go deeper

Related to this question