DSTC6 End to End Conversation Modeling

DSTC6: End-to-End Conversation Modeling Track

Registration

Please register: https://goo.gl/forms/Fxy061gHuSOZGC1i2

News

• Evaluation analysis package: Jan 19 2018

  The package includes all references generated by 11 humans, hypotheses of 20 systems, and evaluation results in DSTC6 end-to-end conversation modeling track. https://www.dropbox.com/s/oh1trbos0tjzn7t/dstc6_t2_evaluation.tgz

• Download the official training data: Sep 7-18 2017

• Test data distribution: Sep 25 2017

• Submission: Oct 8 2017

Track Description

1. Main task (mandatory): Customer service dialog using Twitter

   (*) The tools to download the twitter data and transform to the dialog format from the data are provided.

   Task A: Full or part of the training data will be used to train conversation models.

   Task B: Any open data, e.g. from web, are available as external knowledge to generate informative sentences. But they should not overlap with the training, validation and test data provided by organizers.

2. Pilot task: Movie scenario dialog using OpenSubtitle

• Please cite the following paper if you will publish the results using this setup:

  https://arxiv.org/pdf/1706.07440.pdf

  @article{DSTC6_End-to-End_Conversation_Modeling,
    Author = {Chiori Hori and Takaaki Hori},
    Title = {End-to-end Conversation Modeling Track in DSTC6},    
    Journal = {arXiv:1706.07440},    
    Year = {2017}
  }
  

Necessary steps

Preparation

Most tools are written in python, which were tested on python2.7.6+ and python3.4.1+, and some bash scripts are also used to execute those tools.

For data preparation, you will need additional python modules as follows:

• six
• tqdm
• nltk

which can be installed by

pip install 

or

pip install  -t 

where is a directory storing python modules and needs to be accessible from python, e.g. by including it in PYTHONPATH environment variable.

If you try the baseline system, you will need Chainer http://chainer.org ,a deep learning toolkit, to perform training and evaluation of neural conversation models. Please follow the instruction in ChatbotBaseline/README.md.

Twitter task

1. prepare data set using collect_twitter_dialogs scripts.

   $ cd collect_twitter_dialogs
   $ collect.sh
   

   (a twitter account and access keys are necessary to run the script. follow the instruction in collect_twitter_dialogs/README.md)

2. extract training, development and test sets from stored twitter dialog data

   $ cd ../tasks/twitter
   $ make_trial_data.sh
   

   Note: the extracted data are trial data at this moment.

3. run baseline system (optional)

   $ cd ../../ChatbotBaseline/egs/twitter
   $ run.sh
   

   (see ChatbotBaseline/README.md)

OpenSubtitles task

1. download OpenSubtitles2016 data

   $ cd tasks/opensubs
   $ wget http://opus.lingfil.uu.se/download.php?f=OpenSubtitles2016/en.tar.gz
   $ tar zxvf en.tar.gz
   

2. extract training, development and test sets from stored subtitle data

   $ make_trial_data.sh
   

   Note: the extracted data are trial data at this moment.

3. run baseline system (optional)

   $ cd ../../ChatbotBaseline/egs/opensubs
   $ run.sh
   

   (see ChatbotBaseline/README.md)

Directories and files

• README.md : this file
• tasks : data preparation for each subtask
• collect_twitter_dialogs : scripts to collect twitter data
• ChatbotBaseline : a neural conversation model baseline system

Contact Information

You can get the latest updates and participate in discussions on DSTC mailing list

To join the mailing list, send an email to: ([email protected]) putting "subscribe DSTC" in the body of the message (without the quotes). To post a message, send your message to: ([email protected]).