Bi-Model-Intent-And-Slot

Pytorch implementation of paper "A Bi-model based RNN Semantic Frame Parsing Model for Intent Detection and Slot Filling".

We only build the model with decoder.

We do not tune the hyper-parameters carefully as it is so boring. Obtaining best result of intent accuracy is 0.9843 and f1 score of slot filling is 0.9563 when model runs a lot of epoch(need some lucky), but still lower than the claimed result of that paper(0.9899, 0.9689).

---	Intent acc	Slot filling F1
paper	0.9899	0.9689
reproduce	0.9843	0.9600

Setup

Pytorch>=0.4.0, python3.

python train.py

Reference

1. Dataset and codes calculator F1 score from here.

2. Tricks of dealing with atis dataset from here.