Resolve NLP Bugs

- [ ] Fix informal language indicator.
- Right now the function for this indicator is actually looking for a list of features that appear more often in informal language, but each feature is equally weighted, and all detections are happening at the word level.
- We need to add other features (e.g. sentence fragments and some slang elements like like used as an adjective or interjection),
- It would also be better if we had weights for informality, so that words like THIS or THAT were less heavly weighted than like, like. To do this, we'd have to train a classifier on a labeled corpus. Good job for a grad student sometime?
- in many cases, it would be better to tag entire sentences rather than words. This might be the easiest rewrite of the current indicator. If we added a layer to the indicator so that it returned offsets for whole sentences only if the proportion of informal words was above a threshold, we'd get more plausible results for highlighting from teachers' points of view for relatively small effort.
- [ ] Fix transition words. Right now the temporal transition word indicator is looking to temporal noun phrases but placing no limits on the length or syntactic position of those noun phrases. It should be restricted (a) to limit the length of the noun phrases it recognizes, and (b) to exclude phrases that are functioning as predicates (spacy dependencies like attr)

Locate the informal-language indicator and temporal transition-word indicator, then read how they detect words, noun phrases, offsets, and spaCy dependency roles. Separate the two bug areas before deciding scope. Done means the indicators account for the requested sentence-level and phrase restrictions and have observable behavior covering the described cases.

Written by the indexing model from the issue text.