search

liangeric / nlpQ-A

NLP project that is a Question and Answer system.
2 stars 0 forks source link

Best Answer #7

Closed liangeric closed 3 years ago

liangeric commented 3 years ago

Instead of choosing top 3, maybe keep looping until BERT finds an answer? Seems like from our test cases if a sentence does not have a good answer it returns [CLS], so we can keep looping until something is found, although more testing is needed with extra questions to see if this is a good enhancement.

liangeric commented 3 years ago

On second thought, probably best to still have a threshold, maybe top 5 or 10 (maybe 10 is too much?) sentences to loop through to get possibly have BERT find an answer. If we keep going until an answer is found then we might be returning an answer when in reality an answer can not be found.

raymond-yang commented 3 years ago

I think its reasonable to do a top 5? We just need a way to know if the BERT returned answer is the true answer or as close as we can get to it. Also, still need to keep the potential of unanswerable questions.

liangeric commented 3 years ago

Solution we decided upon was an ensemble method to pick better sentences, and upgrading our BERT model to train on a larger wikipedia dataset, so that answers produced were more likely to appear in top 3.