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
• 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

1. Train/Test: word sequences with IOB slot tags and the indicator of the dialogue start point (1: starting point; 0: otherwise) data/cortana.communication.5.[train/dev/test].iob

Getting Started

You can train and test JointSLU with the following commands:

  git clone --recursive https://github.com/yvchen/ContextualSLU.git
  cd ContextualSLU

You can run a sample tutorial with this command:

  bash script/run_sample.sh memn2n-c-gru theano 0 | sh

Then you can see the predicted result in sample/rnn+emb_H-100_O-adam_A-tanh_WR-embedding.test.3.

Model Running

To reproduce the work described in the paper. You can run the baseline slot filling w/o contextual information using GRU by:

  bash script/run_sample.sh gru theano 0 | sh

Contact

Yun-Nung (Vivian) Chen, y.v.chen@ieee.org

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}
}