Based on the current implementation, the abnormality_score seems like never get passed further to the token_score calculation, while being computed. This makes the score calculation more like the "one-class" design in Sec4.1 in the paper. Could this be the reason for the discrepancy in results?
Based on the current implementation, the abnormality_score seems like never get passed further to the token_score calculation, while being computed. This makes the score calculation more like the "one-class" design in Sec4.1 in the paper. Could this be the reason for the discrepancy in results?