Use sentence-data in database to supplement existing ML model training data

[jecarr]

When a report is submitted, we either (solely) retrieve existing ML models or build/save/retrieve them.

• In report-analysis (training data is self.json_tech)
• Looping through all the techniques and building a model for each one

Our training data is currently what was provided from the initial TRAM repo

• The attack.json training data (saved as self.json_tech in rest_svc.py)
• Loading it in rest_svc.py

It would be good to utilise the true positives, false negatives, and false positives we have in the database as training data

• True Positives are added when we add an attack that was initially predicted as an attack
• False Negatives are added when we add an attack that was not initially predicted as an attack
• False Positives are added when we reject an attack that was initially predicted as an attack
• (We currently do not save True Negatives)

Only add these to the training data if they are associated with sentences from completed reports.

If #88 has been completed, please also exclude sentences with non-confident mappings from the training data.