Optimal hyper-parameters

[thinhphp]

Hi authors,

I couldn't find the optimal parameters to reproduce your results reported in the paper. In detail, I validate your provided model, the results are 84.3%, 72.8% and 38.4% for Slot F1, Intent accuracy and Overall accuracy, respectively. It is a large margin compared to your reported results in the paper (88.3%, 76.3% and 43.5%).
Could you provide more details and/or your pretrained model with the best result?

Thank you

[yizhen20133868]

@phoaiphuthinh Thanks for your interest in our work. Due to some stochastic factors, it is necessary to slightly tune the hyper-parameters using grid search. In our experiments, we carefully tune the hyper-parameters including dropout/batch-size/hidden-size. In addition, in pre-trained models, we carefully tune the hyper-parameters learning-rate (such as 110^-16 ~ 910^-6). Hope it is helpful for you.

[thinhphp]

@yizhen20133868 Thank you for replying. Could you provide the ranges you used to perform grid search for me to reproduce it, or suggest a combination of hyper-parameters that produce comparable results?

[awake020]

For the MixATIS dataset, the hyperparameter settings we have tried include: word_embedding_dim:[64, 128]; intent_embedding_dim[64, 128]; decoder_gat_hidden_dim[64, 128], slot_graph_ window[2, 3, 4]
For the MixSNIPS dataset, the main hyperparameters we try to set are: word_embedding_dim:[64, 128]; intent_embedding_dim[64, 128]; decoder_gat_hidden_dim[64, 128], slot_graph_ window[1]. Thank you for your interest in this work and hope it will help you.

[thinhphp]

Thank you for your reply.