SDEN-Pytorch

Pytorch implementation of Sequential Dialogue Context Modeling for Spoken Language Understanding

Requirements

pytorch==0.4
nltk==3.5.1
sklearn_crfsuite

Run

python3 main.py

Data

I have modified Stanford Multi-turn dataset to fit this model. So it has some noise especially slot tags. It consists of three domain, Weather, Schedule, Navigate. I did dialogue recombination for multi-domain dialogue and modified its format to BIO.

sample

Single domain dialogue

User :  Will it be hot in Inglewood over the next few days?
BOT  :  It will be warm both Monday and Tuesday in Inglewood.
User :  Thank you very much.
BOT  :  You're welcome. Hope you have a great day.

Multi domain dialogue

User :  is it going to be raining this weekend
BOT  :  What city are you inquiring about?
User :  Alhambra please.
BOT  :  It will be raining on Saturday and hailing on Sunday in Alhambra.
User :  Thanks.
BOT  :  happy to help
User :  I need a gas station
BOT  :  I have one gas station listed. Want more info?
User :  What is the address?
BOT  :  76 is at 91 El Camino Real.
User :  Thank you!
BOT  :  You're welcome, stay safe.

Devset Result

Intent Detection : 0.9503091367071216 (Accuracy)

Slot Extraction

	precision	recall	f1-score	support
B-agenda	0.256	0.278	0.267	36
I-agenda	0.733	0.407	0.524	54
B-date	0.826	0.836	0.831	911
I-date	0.533	0.885	0.665	549
B-distance	0.624	0.674	0.648	487
I-distance	0.424	0.353	0.386	167
B-event	0.813	0.793	0.803	517
I-event	0.637	0.847	0.727	367
B-location	0.718	0.928	0.809	572
I-location	0.384	0.950	0.547	280
B-party	0.298	0.807	0.435	187
I-party	0.471	0.471	0.471	17
B-poi_type	0.790	0.738	0.763	534
I-poi_type	0.528	0.718	0.608	301
B-room	1.000	0.400	0.571	35
I-room	0.683	0.848	0.757	33
B-time	0.496	0.595	0.541	220
I-time	0.129	0.286	0.178	14
B-traffic_info	0.661	0.527	0.587	237
I-traffic_info	0.749	0.636	0.688	272
B-weather_attribute	0.904	0.877	0.890	546
I-weather_attribute	0.954	0.775	0.855	80
avg / total	0.683	0.775	0.712	6416