multi-step-training

[WeihaoGe1009]

1. relevance - binary (+NA?)
2. highlight and note - unstructured text, but highlight needs to be exact subtext in selftext
   • condition 1: highlight and note == None and relevance ==False,
   • condition 2: highlight and note != None and relevance == True
   • check highlight be exact subtext (would BWT facilitate it or not?)
3. field - categorical
   • train based on original text ( + masked_author_id?)
   • train based on highlight and notes ( + background?)
4. background - categorical or unstructured text?
5. train conclusion
   • has misunderstanding or has vague knowledge or correct
     • binary (misunderstanding + vague knowledge) vs correct
     • double binary 10, 01, 00 (possible 11?)
   • which misunderstanding/vague knowledge - categorical or unstructured text
     • add a layer/prompt: try put the conclusion in existing categories, if not, create a new category
     • in this case, a train of binary prediction is not suitable. what score and what threshold can figure out the categorical prediction is not working and need to create a new category?
     • a small code of create new category

[WeihaoGe1009]

simplifying: done the simple sentence training first.
Then, we can take out sentence prediction results as an annotation to train for highlighted sentences.