TyDi QA: A Benchmark for Information-Seeking Question Answering in Typologically Diverse Languages

This repository contains information about TyDi QA, code for evaluating results on the dataset, implementations of baseline systems for the dataset, and some advice for working with the dataset.

Want to keep up to date on updates and new releases? Join our low-traffic announcement email list.

Introduction

TyDi QA is a question answering dataset covering 11 typologically diverse languages with 204K question-answer pairs. The languages of TyDi QA are diverse with regard to their typology -- the set of linguistic features that each language expresses -- such that we expect models performing well on this set to generalize across a large number of the languages in the world. It contains language phenomena that would not be found in English-only corpora. To provide a realistic information-seeking task and avoid priming effects, questions are written by people who want to know the answer, but don’t know the answer yet, (unlike SQuAD and its descendents) and the data is collected directly in each language without the use of translation (unlike MLQA and XQuAD).

To see some examples from the dataset with linguistic glosses or for information on TyDi QA's leaderboard, see the website.

For a full description of the dataset, how it was collected, and the quality measurements for the baseline system, see the TACL article.

The Tasks

Primary tasks:
- Passage selection task (SelectP): Given a list of the passages in the article, return either (a) the index of the passage that answers the question or (b) NULL if no such passage exists.
- Minimal answer span task (MinSpan): Given the full text of an article, return one of (a) the start and end byte indices of the minimal span that completely answers the question; (b) YES or NO if the question requires a yes/no answer and we can draw a conclusion from the passage; (c) NULL if it is not possible to produce a minimal answer for this question.
Secondary task:
- Gold passage task (GoldP): Given a passage that is guaranteed to contain the answer, predict the single contiguous span of characters that answers the question. This is more similar to existing reading comprehension datasets (as opposed to the information-seeking task outlined above). This task is constructed with two goals in mind: (1) more directly comparing with prior work and (2) providing a simplified way for researchers to use TyDi QA by providing compatibility with existing code for SQuAD 1.1, XQuAD, and MLQA. Toward these goals, the gold passage task differs from the primary task in several ways:
  - only the gold answer passage is provided rather than the entire Wikipedia article;
  - unanswerable questions have been discarded, similar to MLQA and XQuAD;
  - we evaluate with the SQuAD 1.1 metrics like XQuAD; and
  - Thai and Japanese are removed since the lack of whitespace breaks some tools.

We of course encourage you to participate in the primary tasks as we believe these are a fuller and more robust representative of information-seeking question answering. However, we realize that not all researchers may be able to jump directly into these tasks. If you are constrained by computational resources or are tied to existing code that processes the SQuAD format, the gold passage task may be a better way for you to get started.

When reporting results for any TyDi QA tasks, please include the full task descriptor using one of the strings: TyDiQA-SelectP, TyDiQA-MinSpan, or TyDiQA-GoldP. Please do NOT simply list 'TyDi QA' in your results table, since we do have several flavors of the task, which are quite different from one another and we want to avoid confusion.

Download the Dataset

Once you've chosen which task to work on (above), you can download the data at the following URLs.

For the primary tasks:

The primary task training set is about 1.6GB while the dev set is about 150MB.

For the gold passage task:

The gold passage training set is about 50MB and the dev set is about 10MB. The extra tarball for the dev set contains JSON files that are split along language boundaries; these are used for evaluation while the single large JSON dev file makes it easier to run inference on the entire dev set in a single invocation.

Building a Baseline System

Primary Tasks (TyDiQA-SelectP and TyDiQA-MinSpan)

We provide a baseline system based on multilingual BERT in this repo. Please see baseline/README.md for details on running and modifying that system. You may also find this code useful even if you plan to build a system from scratch as it is designed to be easily re-used.

Gold Passage Task (TyDiQA-GoldP)

Because the gold passage task has been simplified to fit the constraints of the SQuAD 1.1 setting, it can generally be swapped into any code that accepts SQuAD 1.1 JSON inputs by simply changing a few file paths in your code. We provide an example of doing exactly this with the original/unmodified multilingual BERT reference implementation. See gold_passage_baseline/README.md for details.

Evaluation

Primary Task Evaluation

The predictions can be evaluated using a command like the following:

python3 tydi_eval.py \
  --gold_path=tydiqa-v1.0-dev.jsonl.gz \
  --predictions_path=your_model_predictions.jsonl

This script computes language-wise F1 scores and then averages over languages, excluding English. Spans are compared based on predicted byte positions and partial credit is assigned within spans based on F1 positional overlap. See the description of evaluation in the TACL article for details.

Please see the evaluation script for a description of the prediction format that your model should output.

Gold Passage Task Evaluation

For the gold passage task, we re-use the existing SQuAD 1.1 evaluation code to allow maximal re-use of existing pipelines. An example of calling the code for evaluation is in gold_passage_baseline/eval_gold_passage_baseline.sh.

cd gold_passage_baseline
vim eval_gold_passage_baseline.sh  # Edit path to `TYDIQA_GOLDP_DIR`
./eval_gold_passage_baseline.sh predictions.jsonl /tmp

Note that for dev and test evaluation, each language is evaluated separately and the overall score is the average over languages, excluding English.

Leaderboard Submissions

In addition to reporting results on the dev set in your own research articles, we also encourage you to submit to our public leaderboard, to create a record of your experiments. We believe leaderboard submissions serve two main purposes:

(a) to create an existence proof that such a result is possible under carefully isolated conditions (i.e. cheating, intentional or accidental is difficult) so that the community knows such a score is possible; and (b) to inform the community how the result was obtained. Toward this latter goal, we request that you submit a description (e.g. paper draft) of your submission and also answer a few "repoducibility questions" that let the community know if it will be possible to reproduce and build on your result. These include:

Is there a research paper describing the system you are submitting? (The community benefits far more from knowing how to achieve a result than the fact that it exists.)
Is the source code for the system you are submitting publicly available? (Your results will be replicated and trusted more if the community can quickly and reliably reproduce your results).
Was the system you are submitting trained on any additional public data?
Was the system you are submitting trained on any NON-public data? (The community cannot reproduce results on non-public data.)
Was the system you are submitting trained with, or does it use, any external APIs, data labelers, or data transformations (e.g. a translation API)? (The use of public APIs is not reproducible and creates a black box effect since the community does not know the details of the underlying model and data it was built on.)

For step-by-step instructions on submitting, see leaderboard.md.

In addition to submitting to the leaderboard we encourage you to make both your source code and your Docker images public so that others can easily run inference with your system. This opens up the possibility of others (such as MT-focused researchers) building on top of (and citing!) your QA system.

Analyze Your Results

We encourage those working with the data to not only report numeric results, but also analyze the results at a linguistic level. Consider partnering with linguists and/or native speakers of these languages to create glosses that explain how your model is interacting with language. See the TACL article for examples of glossed examples with explanations (Figures 2 - 7).

Source Data

The articles for TyDi QA are drawn from single coherent snapshots of Wikipedia from the Internet Archive to enable open-retrieval experiments. You can download the original article data in Wikitext format from the following URLS:

Podcast

We spoke with AI2's NLP Highlights Podcast about TyDi QA. Have a listen on https://soundcloud.com/nlp-highlights/111-typologically-diverse-multi-lingual-information-seeking-questions-with-jon-clark.

Citation

Please cite the TyDi QA TACL article as:

@article{tydiqa,
title   = {TyDi QA: A Benchmark for Information-Seeking Question Answering in Typologically Diverse Languages},
author  = {Jonathan H. Clark and Eunsol Choi and Michael Collins and Dan Garrette and Tom Kwiatkowski and Vitaly Nikolaev and Jennimaria Palomaki}
year    = {2020},
journal = {Transactions of the Association for Computational Linguistics}
}

Contact us

If you have a technical question regarding the dataset, code or publication, please create an issue in this repository. This is the fastest way to reach us.

google-research-datasets/tydiqa

google-research-datasets

Reviews

Repository Details