use slot-gated model for joint Slot Filling and Intent Prediction. and add CRF layer.

slot-gete论文解读

https://blog.csdn.net/shine19930820/article/details/83052300

CRF

crf开关：--use_crf

预训练

预训练词向量加入之后，效果反而变差了，Semantic Acc: 27.40左右。具体原因不明，有了解过的请告知，感谢。

后期加入bert实验。

实验结果

ATIS数据集	crf	softmax
Slot f1	val: 97.39 ; test: 95.15	val: 97.0; test:95.08
Intent Acc	val: 98.2; test: 95.41	val: 97.6; test: 95.52
Semantic Acc	val: 90.0; test: 83.42	val: 89.0 ; test: 83.53

Environment

tensorflow 1.11 python 3.5

the origin code : https://github.com/MiuLab/SlotGated-SLU

Reference

1. Slot-Gated Modeling for Joint Slot Filling and Intent Prediction : http://www.aclweb.org/anthology/N18-2118