M2QA (Multi-domain Multilingual Question Answering) is an extractive question answering benchmark for evaluating joint language and domain transfer. M2QA includes 13,500 SQuAD 2.0-style question-answer instances in German, Turkish, and Chinese for the domains of product reviews, news, and creative writing.
This repository accompanies our paper "M2QA: Multi-domain Multilingual Question Answering" and provides access to the benchmark dataset, our custom-built annotation platform, and the code to reproduce all our experiments.
The M2QA benchmark dataset consists of 13,500 SQuAD 2.0-style question-answer instances, divided evenly across nine language-domain combination pairs (1500 instances each). 40% of the data are unanswerable questions, 60% are answerable.
The data is provided in m2qa_dataset/ and uploaded on Hugging Face: https://huggingface.co/datasets/UKPLab/m2qa
Following Jacovi et al. (2023), we encrypt the data to prevent leakage of the dataset into LLM training datasets. Have a look at m2qa_dataset/README.md to see how easily you can use the dataset.
We also provide additional training data for five domain-language pairs, consisting of 1500 question-answer instances each, totalling 7500 training examples. This data is available in the m2qa_dataset/Additional_Training_data directory.
Our main contributions include:
- Baselines: We evaluate baseline and transfer performance on M2QA using a wide range of models and transfer techniques, including fully-finetuned models, modular transfer learning and LLMs.
- Domain and language are not independent axes: We find that transfer performance considerably varies across domain-language combinations.
- SQuAD 2.0 metric not applicable in every language: We find that the widely used SQuAD 2.0 evaluation metric is insufficient for evaluating multilingual extractive QA due to its reliance upon whitespace tokenization and propose a version of the metric that mitigates the issue. For more information, have a look at Section 5.1 of our paper or Experiments/M2QA_Metric/README.md.
- More research has to be done on joint language and domain transfer! Our results show that modern LLMs perform considerably worse on their target than on their source domain-language pair, highlighting the need for further research into methods that transfer both linguistic and domain-specific information.
We provide the code to reproduce all our experiments in the Experiments/ directory. Have a look at Experiments/README.md for a detailed explanation of our experiments and how to reproduce them. There, you find everything, including the code to reproduce these main results:
The M2QA data in Chinese "Creative Writing" and "Product Reviews" do not contain whitespaces. This leads to reduced performance of XLM-RoBERTa. We find that the performance of XLM-RoBERTa can be drastically increased by simply adding whitespaces between every word:
We have developed a new annotation platform to fulfil all of our requirements. The source code can be found in the Website/ directory. In the Website/README.md, you can find screenshots of the website and instructions on how to set it up.
If you use M2QA in your work, please consider citing our paper:
@article{englaender-etal-2024-m2qa,
title="M2QA: Multi-domain Multilingual Question Answering",
author={Engl{\"a}nder, Leon and
Sterz, Hannah and
Poth, Clifton and
Pfeiffer, Jonas and
Kuznetsov, Ilia and
Gurevych, Iryna},
journal={arXiv preprint},
url="https://arxiv.org/abs/2407.01091",
month = jul,
year="2024"
}
The M2QA dataset is distributed under the CC-BY-ND 4.0 license.
Following Jacovi et al. (2023), we decided to publish with a "No Derivatives" license to mitigate the risk of data contamination of crawled training datasets.
The code of this repository (i.e. everything except the M2QA dataset) is distributed with the Apache License 2.0.
Contact persons:
- Leon Engländer: 📧 Email | 🐦 Twitter | 💻 GitHub
- Hannah Sterz: 📧 Email | 🐦 Twitter | 💻 GitHub
- Ilia Kuznetsov: 📧 Email | 🐦 Twitter | 💻 GitHub
If you have any questions, please do not hesitate to contact us or (preferably) open an issue here on GitHub.
https://www.ukp.tu-darmstadt.de/
UKP Lab is part of the TU Darmstadt: https://www.tu-darmstadt.de/
This repository contains experimental software and is published for the sole purpose of giving additional background details on the respective publication.