/converse_reading_cmr

code, data, and model for ACL-19 paper "Conversing by Reading: Contentful Neural Conversation with On-demand Machine Reading"

Primary LanguagePythonMIT LicenseMIT

Code for "Conversing by Reading: Contentful Neural Conversation with On-demand Machine Reading"

This is the code for the following paper:

Conversing by Reading: Contentful Neural Conversation with On-demand Machine Reading
Lianhui Qin, Michel Galley, Chris Brockett, Xiaodong Liu, Xiang Gao, Bill Dolan, Yejin Choi, Jianfeng Gao; ACL 2019

This package contains two independent codebases for (1) re-creating the dataset of our experiments and (2) the CMR model described in the paper. All this code is made publicly and freely available with the final version of our paper and will be hosted on github.

Data

You can download the raw data from here directly (then you can just skip below steps).

For the purpose of reproducbility, we also provide code to recreate our dataset (note that both running this code or downloading the above data will yield the same output). Since the raw sources to recreate the data are static (Reddit and Common Crawl dumps), this ensures the data output remains the same, making our experiments reproducible.

How to run: After moving into the data, data extraction consists of a single command (make -j4), but the README file gives details about software and packages to install and further information about the data.

Notes:

  • The full data extraction pipeline may take 1-5 days, depending on compute power and internet speed;
  • In some rare cases, data extraction output might have slight differences across runs (< 0.1% of the data) due to 503 errors returned by the Common Crawl server. The final version will better handle these rare cases.

Evaluation

The test data and evaluation scripts are under evaluation/. The README file gives the details.

Model

We provide code to train and test with the CMR model, which is described in details in this README.

How do I cite Conversing by Reading?

@inproceedings{qin-etal-2019-conversing,
    title = "Conversing by Reading: Contentful Neural Conversation with On-demand Machine Reading",
    author = "Qin, Lianhui and Galley, Michel and Brockett, Chris and Liu, Xiaodong 
              and Gao, Xiang and Dolan, Bill and Choi, Yejin and Gao, Jianfeng",
    booktitle = "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
    month = "jul",
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/P19-1539",
    pages = "5427--5436",
}