Azure ML

After bootstrapping a dead-simple language model for task extraction using Microsoft LUIS (Language Understanding Intelligent Service)¹, then building a prototype task service and OneNote add-in around it, I wanted to extend the model significantly. My use of LUIS did not really fit into its prescribed “utterances—entities—intents” taxonomy; further, LUIS is not extensible enough for the kind of custom language modeling I wanted to do.

Enter: Azure Machine Learning. Azure ML is a newly-revamped² suite of machine learning services and tools. It includes a Workbench for data preparation and model development using Python and family (Anaconda, Jupyter, scikit-learn, spaCy… whatever you need); an Experimentation service for running experiments, locally or remotely, on your local machine or Docker or Spark; and a Model Management service for deploying models behind a web service.

So after iterating on an NLP model for task classification (read all about it here), I turned to Azure ML to deploy it so that I could swap out the API call to LUIS with an API call to this new service hosting my custom NLP model³.

Azure ML is still in preview, and so there were various hiccups in getting my model deployed⁴. However, the team was very responsive to my questions and bug reports, and I eventually got things up and running. Because of this active engagement, I’m confident stability will continue to improve during preview.

Prototype of OneNote Task Extraction add-in with toy example (GIF)

Prebuilt vs Custom ML

My experience with using both LUIS and Azure ML has illustrated to me the importance of having both prebuilt/API-like and custom ML solutions available for developer use. LUIS as a prebuilt solution was excellent for bootstrapping my idea and getting something off the ground. Once I felt more comfortable in the space and had the context to apply more domain-specific ideas, I could easily transition to Azure ML, spaCy, scikit-learn, etc. to implement them.

Footnotes