David Flanagan [:djf] Assignee
	Description • 11 years ago

For speed, the prediction engine does a fairly quick and constrained search for corrections. It has been refactored to be interruptable, however, so when it fails to find any suggestions, we might as well go ahead and ask it to try again, but look harder this time.

	David Flanagan [:djf] Assignee
	Comment 1 • 11 years ago

I'm removing this from the list of bugs blocking bug 797170. I'll create the simple patch described here if, in testing, we find that there are words where the desired suggestion is not coming out correctly.

	David Flanagan [:djf] Assignee
	Comment 2 • 11 years ago

This patch modifies worker.js so that if the first call to predict() doesn't return any suggetions, it tries again, but search for words with edit distance 2 instead of edit distance 1.

There are chagned in predictions.js to make predict() configurable this way.

The results are really good.  We can now make corrections like caebations->carnations and teatinf->testing.  The initial search for caebations takes ~20ms, and the deeper search takes 120ms.  The initial search for teating takes 24ms and the deeper search takes 86ms.  So I think this new strategy of a shallow search followed by a deep search totally wins.

I wish I'd done it a long time ago!

	David Flanagan [:djf] Assignee
	Comment 3 • 11 years ago

Landed on master https://github.com/mozilla-b2g/gaia/commit/0225f1ac137f619043054a7b05c998b89fbfaaaf

	David Flanagan [:djf] Assignee
	Comment 4 • 11 years ago

Uplifted to v1-train in bug 873934