1.is there an attention on Intent Detection task in your code? you seem to just make a full-connected layer after Bi-LSTM. That confused me.
2.why the final intent vector is encoder_final_state_h rather than summation of all encoder_state_h_i?
Yes you are right. I didn't use attention in the classification model. I would add this part or you can commit your code if you like. Thank you for pointing this out.
1.is there an attention on Intent Detection task in your code? you seem to just make a full-connected layer after Bi-LSTM. That confused me.
2.why the final intent vector is encoder_final_state_h rather than summation of all encoder_state_h_i?