Good training but bad results

[iuria21]

Hello, first thanks for your work. I trained the model with the same data as you, and I got similar results in the training (except the time):

But then when I try to correct some examples the result is awful... these are the outputs for your same examples:

Do you know why could it be? The vocabulary is the same, and I don't know why the number of sentences it's not the same but very similar. And, do you think that it will succes if i feed it with more data? I also tried with spanish and the result is not satisfactory...

Thanks!

[shibing624]

good question.

[HariWu1995]

I think that it's not enough noise in the train set. One suggestion is that you increase the threshold one by one for train your model to work with difficult sentences.

[shibing624]

try seq2seq with attention, your testing loss is 1.248 is not good enough. i think 0.05 is good for this task. 55873 281/284 [============================>.] - ETA: 6s - loss: 0.0510 55874 282/284 [============================>.] - ETA: 4s - loss: 0.0510 55875 283/284 [============================>.] - ETA: 2s - loss: 0.0510 55876 284/284 [==============================] - 599s 2s/step - loss: 0.0510 - val_loss: 0.2124

[SuMyo]

@shibing624 could you please share the code sample for using seq2seq attention?

[shibing624]

https://github.com/shibing624/pycorrector