Training.
Ideally, we would train our classifiers on actual examples of web API request URLs and the corresponding descriptions, but such data is not easily available.
We instead created a set of URLs by manually navigating through the top 10 US websites and recording all HTTP requests made by the browser.
The resulting data set contains 2200 unique URLs from 234 different domains.
While they do not all reflect API calls, we believe they are sufficiently similar in structure for our training to be effective.
We trained classifiers for 5 parameter tags, manually marking (overall) 2300 positive and negative examples.