Prepare non-English SQuAD

[yingzwang]

I saw that you released GermanQuAD. (https://huggingface.co/datasets/deepset/germanquad)
Could you share some details on how you prepared the dataset?
I want to train a monolingual Dutch QA model. I plan to use an existing pre-trained Dutch BERT base model, and prepare some Dutch SQuAD samples myself. I have following ideas,

1. machine translation of English SQuAD.
2. human annotation of 500 Dutch question-passage-answer (Q-P-A) triplets in my domain.

If I skip 2), would 1) already be enough?
If I add 2), would 500 Q-P-A samples be enough?
By enough I mean at least better than a multilingual model.

I've tried the multilingual model deepset/xlm-roberta-large-squad2 on my Dutch samples. It works a bit however the performance is not satisfactory, which is expected. I feel that training a monolingual Dutch QA model will be much better. Your advice and tips are much appreciated!

[bogdankostic]

Hi @yingzwang! We released a paper where we summarize our learnings from creating the GermanQuAD dataset. We saw that using human annotations instead of machine-translated labels results in a better performance and also did some experiments on how the number of training samples affects the performance (see Figure 3).
Also, we released a blogpost about QA on different languages than English.

Let me know if you have further questions :)

[yingzwang]

Thanks for sharing the paper @bogdankostic ! At a quick glance I see a lot of useful stuff that's exactly what I'm looking for.
The result of Table 4 is a bit surprising to me. The monolingual model trained on translated SQuAD is even worse than the XLM model trained on English SQuAD! Is the machine translated SQuAD that bad? I manually checked a few machine translated (English-to-Dutch) SQuAD samples, and they look pretty OK, i.e., the translated text is quite accurate and natural. What could seriously go wrong with the machine translated SQuAD, so that the monolingual model cannot learn well?

The performance boost by human annotated GermanQuAD is significant. I'm encouraged to invest more effort preparing more human annotated DutchQuAD :) Also need to explore a bit your annotation tool.

[yingzwang]

Also in the paper you mentioned about "detailed labeling instructions". Could you share your complete labeling instructions? Or is it already everything in Appendix A?

[julian-risch]

Hi @yingzwang Appendix A contains all labeling instructions, yes. As we had a team of annotators, we did regular meetings during the annotation process where we discussed newly annotated samples. For example, we discussed the phrasing of the question and whether we can rephrase the question so that there is less lexical overlap with the answer string. Another example is that we discussed whether the question is self-sufficient or whether there should be some additional information in the question for open-domain question answering.

English-to-Dutch machine translation of SQuAD should already give you quite good results. At first glance, the translations look very good, I agree. However, sometimes there are small mistakes or the words used in the translation do not really fit the context. Sometimes machine-translation could increase the lexical overlap of the question and the answer and thus simplify the task for the model. That might be one reason why training on machine-translated data is not as good as training on carefully hand-annotated data.

I think 500 hand-annotated samples could already be enough to see an improvement, yes. I would recommend that you try to make these 500 samples not too simple for the model.

Please keep us updated about the progress of your project and don't hesitate to contact us again if there are any questions coming up. Feel free to close the issue for now if there are no further questions at the moment and re-open it later. Good luck with training a monolingual Dutch QA model! 👍

[yingzwang]

Thanks for the elaborated answer @julian-risch ! You have really adopted a robust quality control workflow. How large is your annotation team? Not sure if I can afford a similar setup but I'll try.

I found the manual of your annotation tool here. There's one rule saying "Don’t create question that elaborate on context", in page 13. Why is that? A user is very likely to ask a question in that way, especially in a conversation. Why should we avoid it in training data?

Your labeling instructions are insightful. From ~50 annotated samples I've collected so far, some already violated the "no lexical overlap rule" and "prefer short answer over long answer rule". I read your paper just in time :)

[julian-risch]

Great to hear that you find our documentation helpful! @yingzwang
Questions shouldn't have too many words in common with the context around the answer because that simplifies the question answering task and simple questions are not very helpful for the training. You are right, that users might ask such questions in an application. That's no problem for the model as it is a simple task to then find the corresponding document and answer. There is no need for training data that represents this relatively simple task.
We had a team of five annotators when we created the GermanQuAD dataset. Having a team of annotators does not only speed up the annotation process. It also helps at formulating more diverse questions and answers. For that reason I would recommend to have at least two annotators.

[aantti]

There's also this recent article we did on annotation https://www.deepset.ai/blog/labeling-data-with-haystack-annotation-tool (and the latest one is about evaluation). Just in case it might be helpful in addition to the documentation pages.