Maluuba NewsQA

Tools for using Maluuba's news questions and answer data. The code in the repo is used to compile the dataset since it cannot be made directly available due to legal reasons.

You can find more information with stats on the dataset here.

Data Description

We originally only compiled to CSV but now we also build a JSON file.

JSON

Here is an example of how the data in combined-newsqa-data-v1.json will look:

{
    "data": [
        {
            "storyId": "./contoso/stories/2e1d4",
            "text": "Hyrule (Contoso) -- Tingle, Tingle! Kooloo-Limpah! ...These are the magic words that Tingle created himself. Don't steal them!",
            "type": "train",
            "questions": [
                {
                    "q": "What should you not do with Tingle's magic words?",
                    "consensus": {
                        "s": 115,
                        "e": 125
                    },
                    "isAnswerAbsent": 0.25,
                    "isQuestionBad": 0.25,
                    "answers": [
                        {
                            "sourcerAnswers": [
                                {
                                    "s": 115,
                                    "e": 125
                                }
                            ]
                        },
                        {
                            "sourcerAnswers": [
                                {
                                    "noAnswer": true
                                }
                            ]
                        }
                    ],
                    "validatedAnswers": [
                        {
                            "s": 115,
                            "e": 125,
                            "count": 2
                        },
                        {
                            "noAnswer": true,
                            "count": 1
                        },
                        {
                            "badQuestion": true,
                            "count": 1
                        }
                    ]
                }
            ]
        }
    ],
    "version": "1"
}

Explanation:

Explanation of the answer fields:

CSV

This section describes the CSV formatted files for the dataset. We originally only compiled to CSV.

The combined dataset in the .csv file is made of several columns to show the story text and the derived answers from several crowdsourcers.

There are some other fields in combined-newsqa-data-*.csv for raw data collected when crowdsourcing such as the validation of collected data.

Requirements

Run either the Docker steps or do the manual set up.

The dataset for character based indices will be the combined-newsqa-data-*.csv file in the root.

The dataset for token based indices will be maluuba/newsqa/newsqa-data-tokenized-*.csv.

(Recommended) Docker Set Up

These steps handle packaging the dataset and running the tests.

• Clone this repo.
• Download the tar.gz file for the questions and answers from here to the maluuba/newsqa folder. No need to extract anything.
• Download the CNN stories from here to the maluuba/newsqa folder (for legal and technical reasons, we can't distribute this to you).
• In the root of this repo, run:

docker build -t maluuba/newsqa .
docker run --rm -it -v ${PWD}:/usr/src/newsqa --name newsqa maluuba/newsqa

You now have the datasets. See combined-newsqa-data-*.json, combined-newsqa-data-*.csv, or maluuba/newsqa/newsqa-data-tokenized-*.csv.

Tokenize and Split

If you want to tokenize and split the data into train, dev, and test, to match the paper run, then you must get "into" the container and run the packaging command:

docker run --rm -it -v ${PWD}:/usr/src/newsqa --name newsqa maluuba/newsqa /bin/bash --login -c 'python maluuba/newsqa/data_generator.py'

The warnings from the tokenizer are normal.

Troubleshooting Docker Set Up

If you run into issues such as the tokenization not unpacking, then you may need to give Docker at least 4GB of memory.

Manual Set Up

• Clone this repo.
• Download the tar.gz file for the questions and answers from here to the maluuba/newsqa folder. No need to extract anything.
• Download the CNN stories from here to the maluuba/newsqa folder (for legal and technical reasons, we can't distribute this to you).
• Use Python 2.7 to package the dataset (Python 2.7 was originally used to handle the stories and they got encoded strangely - once the dataset is packaged by these scripts, you should be able to load the files with whatever tools you'd like). You can create a Conda environment like so:

conda create --name newsqa python=2.7 "pandas>=0.19.2"

• Install the requirements in your environment:

conda activate newsqa && pip install --requirement requirements.txt

• (Optional - Tokenization) To tokenize the data, you must install a JDK (Java Development Kit) so that you can compile and run Java code.
• (Optional - Tokenization) To tokenize the data, you must get some JAR files. We use some libraries from Stanford. You just need to put the English option of version 3.6.0 in the maluuba/newsqa folder.

Package the Dataset

Tokenize and Split

To tokenize and split the dataset into train, dev, and test, to match the paper run:

python maluuba/newsqa/data_generator.py

The warnings from the tokenizer are normal.

Testing

To make sure that everything is extracted right, run

python -m unittest discover .

All tests should pass.

PEP8

The code in this repository complies with PEP8 standards with a maximum line length of 99 characters.

Legal

Notice: CNN articles are used here by permission from The Cable News Network (CNN). CNN does not waive any rights of ownership in its articles and materials. CNN is not a partner of, nor does it endorse, Maluuba or its activities.

Terms: See LICENSE.txt.