So what I’ve one is identified words the statistical model doesn’t work well for (it’s/its, a/an, there/their, etc), moved them away from the misused word list, and created grammar rules for them.

This has given me some wiggle room to cut my biasing down. I’m aiming for a 0.0050 false positive rate–it’s at 0.0011 and 0.0029% in my tests now. Once I change this balance you’ll notice the misused word detection is more accurate.

What makes this problem challenging is for some words (depending on the context) either word could feasibly be correct. The solution is to look at more context but the more context I choose to use the more danger I have of data sparseness leading to more false positives and less coverage of how the words are actually used. It’s an interesting problem.

What is nice (from my perspective) is where one tool falls short, I have another tool I can use to make up the difference (such as the grammar checker).