Closed HongShihan closed 1 year ago
Hi, Sorry for the late reply! Your assumption make sense because we compute the entity probability by taking softmax over all candidates. In this paper, we tune the threshold for my specific number of candidates. So we may need to use the dynamic threshold or tune a more suitable threshold for more candidates. Also, we note that you can also choose not to use the re-rank probability to compute the mention probability, which we found doesn't hurt the performance and prevents the issue you mentioned here.
Hello, thank you for your very interesting work. I noticed that the final probability of the detected entities tends to decrease when i increase the number of candidates, which sometimes causes the result to be less accurate with increased k as the probability falls below the threshold. Is this normal behavior? Possibly due to the re-ranking function?