Home

NEURAL NETWORKS FOR DIALOG SYSTEMS

DIALOG STATE TRACKING

In contrast to Question Answering systems that answer a question given a context, Dialog Systems need to keep track of the state of the dialogue, acquire evidence from the conversation, perform slot filling and finally issue an api call, which will fetch information from knowledge base (database) or perform an action. Observe in the figure below, a sample dialogue between a virtual agent and a customer, in restaurant domain.

Tasks

1. Issue API call
2. Update API calls
3. Display Options
4. Provide extra information, like address, phone, etc,

Neural Network Architectures

• Jason Weston, Learning End-to-End Goal Oriented Dialog

Research Papers

• Matthew Henderson’s PHD thesis, Discriminative Methods for Statistical Spoken Dialogue Systems
• Machine Learning for Dialog State Tracking : A Review
• Word-based Dialog State Tracking with RNNs
• Deep Neural Network Approach for the Dialog State Tracking Challenge
• Robust Dialog State Tracking Using Delexicalised Recurrent Neural Networks and Unsupervised Adaptation
• DTSC 1, DTSC 2, DTSC 3, DTSC 4, DTSC 5

Note : Download the whole list of papers as zip

Datasets

The (6) dialog bAbI tasks

A set of 6 tasks for dialogue systems in the restaurant domain. Each task tests a unique aspect of dialog. Download from here.

Dialogue State Tracking Challenge Series

• DSTC 1
• DSTC 2
• DSTC 3
• DSTC 4
• DSTC 5

voicy-ai/DialogStateTracking

Setup

git clone https://github.com/voicey-ai/DialogStateTracking
cd DialogStateTracking
# install dependencies
sudo -H pip3 install -r requirements.txt
# train; note : pretrained models are available in repository
#   python3 main.py --train --task_id=1
# interactive mode
python3 main.py -i --task_id=1

Sample Interaction