YUNOVA · Natural Language Processing
Module 10: Production NLP
Deploying an NLP model as an API
Key points
- Wrap the model in a web framework such as FastAPI to serve predictions.
- Load the model once at startup, not on every request.
- Package it with Docker so it runs the same everywhere.
Multilingual NLP for Indian languages
Key points
- Multilingual models such as mBERT, XLM-R and IndicBERT cover many Indian languages.
- Scripts, code-mixing and transliteration make Indian-language text harder.
- Collections such as IndicNLP and AI4Bharat provide datasets and models.
Speech to text with Whisper
Key points
- Whisper is an open speech recognition model from OpenAI.
- It can transcribe many languages and translate speech into English.
- Larger model sizes are more accurate but slower.
Bias and fairness in NLP models
Key points
- Models can learn social biases present in their training data.
- Bias can be measured by comparing outputs across groups.
- Balanced data and careful evaluation help reduce harm.