Tweak question generation

[aolney]

Analyze batch output for common error patterns

[aolney]

Some thoughts:

• Stanford dependencies make the complement of a copula the root of the parse, which created weird predicate-driven questions
• The adjunct questions are often deeply embedded in the sentence, so "in place" wh substitution is easy to overlook and creates working memory problems

Because of the new deep learning work, I'll be seriously developing this code moving forward (it was a placeholder). Towards that end, I'm:

• Removing the adjunct questions until I have movement and aux inversion set up properly
• No longer using the raw dependency parse for question generation
• Focusing on verb frames to create shorter, more directed questions that still work with "in place" wh substitution

[aolney]

Update: using dependency generation as a backup when srl fails

Need better handling of relpro modifiers