Named-Entity-Recognition-with-Bidirectional-LSTM-CNNs
Named-Entity-Recognition-with-Bidirectional-LSTM-CNNs copied to clipboard
kamalkraj
Different understanding about the original paper
In README, you said: “The model produces a test F1_score of 90.9 % with ~70 epochs. The results produced in the paper for the given architecture is 91.14 ” In fact, the paper said the result 91.14 is produced under the situation "All other hyper-parameters and features remain the same as our best model in Table 5", that is , lex feature is used, while you do not use that feature, so this architecture can not reach 91.14.
Correct . I didn't use any lex feature so model can't reach 91.4
On Tue, 12 Jun 2018 at 09:32, lanzhuzhu [email protected] wrote:
In README, you said: “The model produces a test F1_score of 90.9 % with ~70 epochs. The results produced in the paper for the given architecture is 91.14 ” In fact, the paper said the result 91.14 is produced under the situation "All other hyper-parameters and features remain the same as our best model in Table 5", that is , lex feature is used, while you do not use that feature, so this architecture can not reach 91.14.
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/kamalkraj/Named-Entity-Recognition-with-Bidirectional-LSTM-CNNs/issues/4, or mute the thread https://github.com/notifications/unsubscribe-auth/AQTgmg44TgVmr_guDHMjbgd6hjvMBrc8ks5t7z1XgaJpZM4UjvlI .
Did you use Viterbi to do the decoding of the best sequence ? That might also explain the different results
Did you use Viterbi to do the decoding of the best sequence ? That might also explain the different results
If I am not wrong, the code here does not include the transition matrix of the tags. So no need to apply Viterbi here. But this is also a big difference.
