jonze 0.0.20

Reusable Joint Slot and Intent Extraction implementation in Tensorflow2.0

Jonze

Joint Slot and Intent Extraction implementation in Tensorflow2.0

Contains restructured code from the following repo: https://github.com/shubham8111/Joint-NLU

Implementation of Bi-LSTM based NLU baseline and SlotGated-SLU (Goo et al, 2018)(https://www.csie.ntu.edu.tw/~yvchen/doc/NAACL18_SlotGated.pdf) Models are evaulated on Snips and ATIS datasets.

Experiments did not reproduce improvements by SlotGated model over Basline model, on snips dataset.

Preprocessing modules reused from following repo: https://github.com/MiuLab/SlotGated-SLU/

Usage

To install package:
pip install jonze
To train model:
from jonze import train train(dataset = "joint-nlu", datasets_root = "dataset", models_root = "model", layer_size=12)
To test model:
from jonze import test test(dataset = "joint-nlu", datasets_root = "dataset", models_root = "model", layer_size=12, batch_size=46)

Results

Snips Dataset:

Model	Slot F1	Intent accuracy	Semantic Accuracy
Baseline	84.30	96.57	66.43
Slot Gated	83.5	95.57	66.85

Atis Dataset:

Model	Slot F1	Intent accuracy	Semantic Accuracy
Baseline	95.08	94.62	81.97
Slot Gated	94.57	96.41	83.65

P.S. Sometimes Slot F1 might get stuck at zero during training, better weight intialization or training a few epochs only on slot loss can resolve the issue.