Modify the architecture of the SlotGated and get better results.

[jasonchan117]

I did several experiments on it by adding the slot attention context vector to predicting the intent. The slot f1 is 89.05, while the intent accuracy is 97.42 with the same settings from the paper, which is better than the original results. I also did other experiments by removing the intent attention layer and the gate to make some comparisons.

• Execution: Get the results by the below commands.
  • --interplay Add the slot context vectors to the intent prediction.
  • --remove_intent_attn Remove the intent attention.
  • --remove_gate Remove the gate.