ContextualSLU: Multi-Turn Spoken/Natural Language Understanding

A Keras implementation of the models described in [Chen et al. (2016)] (https://www.csie.ntu.edu.tw/~yvchen/doc/IS16_ContextualSLU.pdf).

This model implements a memory network architecture for multi-turn understanding, where the history utterances are encoded as vectors and stored into memory cells for the current utterance's attention to improve slot tagging.

Content

• Requirements
• Dataset
• Getting Started
• Model Running
• Contact
• Reference

Requirements

1. Python
2. Numpy pip install numpy
3. Keras and associated Theano or TensorFlow pip install keras
4. H5py pip install h5py

Dataset

Consists of two datasets derieved from Simulated-Google Restaurant dataset

• Dataset without acts : Only take present utterance intent and slot
• Datset with acts : Take present utterance intent, slot and user_act.

Getting Started

You can train and test JointSLU with the following commands:

  git clone --recursive https://github.com/halloTheCoder/ContextualSLUChatBot.git
  cd ContextualSLUChatBot

Model Running

Points to consider while running

• Give pretrained-embeddings as ConceptNet Numberbatch, other will lead to error.
• Change data directories in sequence_tagger.py, by default dataset without acts folder. (NOTE :: Don't use dataset folder as acts is not handled in code for now)

	python sequence_tagger.py

Contact

• Email - akash.chandra8d@gmail.com

Reference

Main papers to be cited

@Inproceedings{chen2016end,
  author    = {Chen, Yun-Nung and Hakkani-Tur, Dilek and Tur, Gokhan and Gao, Jianfeng and Deng, Li},
  title     = {End-to-End Memory Networks with Knowledge Carryover for Multi-Turn Spoken Language Understanding},
  booktitle = {Proceedings of Interspeech},
  year      = {2016}
}