Filter searches before citation detector

[matt-bernhardt]

This moves the extract_features method in Detector::MlCitation into the initialize method, allowing us to quickly check for whether a given term has enough non-zero features to make it worth calling the detector.

From our analysis in the TACOS notebooks, we believe that phrases which result in only two non-zero values among all their features will never end up being a citation - and that this threshold will allow us to skip the citation detector in 90% of searches.

The filtering is performed in a convenience method named enough_nonzero_values? (naming things is hard).

There is one side effect worth noting: the @Detections instance variable is now defined as false at the top of the initialize method, before the first guard clause, so that we get a consistent Boolean value in all conditions. This required one test to change that previously expected a nil from the guard clause.

I have one concern about this approach, but I'd like to talk about it as part of code review: This approach doesn't seem to leave any traces for us to diagnose after the fact. I thought about whether there should be a data model change here, to indicate which Term records were filtered by this approach, but we haven't talked about that in any detail yet.

Some of the ways that we might use this information would be:

1. To assemble a new batch of training data that has survived the filter;
2. To maintain awareness of how the filter might need to change as we update what features we extract or how the algorithm is trained.

An alternative arrangement to what I'm proposing here would be to have the number of nonzero features calculated in a public method, ideally during feature extraction in Detector::Citation. That might allow that value to be returned via GraphQL for training notebooks, or consulted internally without having to recalculate it. The actual filter could still be internal to Detector::MlCitation, but enough_nonzero_features? would just be an inequality check without the calculation.

Developer

Ticket(s)

https://mitlibraries.atlassian.net/browse/TCO-190

Accessibility

• ANDI or Wave has been run in accordance to our guide and
  all issues introduced by these changes have been resolved or opened
  as new issues (link to those issues in the Pull Request details above)
• There are no accessibility implications to this change

Documentation

• Project documentation has been updated, and yard output previewed
• No documentation changes are needed

ENV

• All new ENV is documented in README.
• All new ENV has been added to Heroku Pipeline, Staging and Prod.
• ENV has not changed.

Stakeholders

• Stakeholder approval has been confirmed
• Stakeholder approval is not needed

Dependencies and migrations

YES dependencies are updated

NO migrations are included

Reviewer

Code

• I have confirmed that the code works as intended.
• Any CodeClimate issues have been fixed or confirmed as
  added technical debt.

Documentation

• The commit message is clear and follows our guidelines
  (not just this pull request message).
• The documentation has been updated or is unnecessary.
• New dependencies are appropriate or there were no changes.

Testing

• There are appropriate tests covering any new functionality.
• No additional test coverage is required.

[matt-bernhardt]

Okay, @JPrevost - I think this is ready for your review again.