I believe I got it working. I think there was something wrong in the way parameters were being ordered in train time vs apply time. We now have a simpler model (only one embedding for the whole chunk of details) vs comparisons for each part vs each other part. I think the field that is messing up the classification the most is the co_authors / co_contributors.
Testing that now but additionally with this new model we should go through and get log proba for each and make sure our annotation is correct. We may need more positive examples too.
I believe I got it working. I think there was something wrong in the way parameters were being ordered in train time vs apply time. We now have a simpler model (only one embedding for the whole chunk of details) vs comparisons for each part vs each other part. I think the field that is messing up the classification the most is the co_authors / co_contributors.
Testing that now but additionally with this new model we should go through and get log proba for each and make sure our annotation is correct. We may need more positive examples too.