Multilingual Question Answering
Multilingual Question Answering (Cross-lingual MRC) · Also known as: cross-lingual question answering, multilingual QA, multilingual MRC, cross-lingual machine reading comprehension
Multilingual question answering (QA) enables a model to read a passage and answer questions in multiple languages, often by fine-tuning a cross-lingual pretrained transformer such as mBERT or XLM-R on an annotated QA dataset in one language and transferring that capability zero-shot or few-shot to other languages. It is the standard approach for building multilingual reading-comprehension and open-domain QA systems.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use multilingual QA when you need to answer questions from passages in several languages and either lack large annotated datasets in all target languages or want a single unified model. It is ideal for multilingual knowledge bases, customer support in diverse regions, and academic research on low-resource languages. Do not use it when queries are confined to a single language — a monolingual model will nearly always outperform a multilingual one due to the capacity trade-off of sharing parameters across languages. Also avoid it when the domain is highly specialised and target-language annotation is affordable; domain-specific fine-tuning on each language separately will yield higher accuracy.
Strengths & limitations
- Enables QA across dozens of languages with a single model, greatly reducing engineering overhead.
- Zero-shot cross-lingual transfer allows deployment in languages with little or no labelled data.
- Strong off-the-shelf performance when built on large pretrained models such as XLM-R Large.
- Standardised benchmarks (MLQA, XQuAD, TyDi QA) support reproducible comparison across systems.
- Both extractive and generative (seq2seq) formulations are available for different answer formats.
- Can be further specialised by continued pretraining on target-language text before fine-tuning.
- Multilingual models trade capacity across languages; per-language performance is typically lower than dedicated monolingual models.
- Low-resource languages with limited pretraining text remain significantly weaker than high-resource ones.
- Computational cost is high: large cross-lingual transformers require GPU resources for both training and inference.
- Cross-lingual transfer degrades for typologically distant languages or scripts absent from the pretraining corpus.
- Evaluation benchmarks often focus on extractive span prediction; open-ended or generative QA across languages is less standardised.
Frequently asked
Which backbone model should I use for multilingual QA?
XLM-R (especially XLM-R Large) is the most widely used backbone due to its strong cross-lingual transfer across 100 languages. For generative or abstractive answers, mT5 is a good choice. If your target languages are European, multilingual BERT may be sufficient and is lighter to fine-tune.
Can I achieve good results with no target-language training data?
Zero-shot transfer from English-only fine-tuning often yields usable results on high-resource languages closely related to English, but performance drops noticeably on typologically distant or low-resource languages. Even a few hundred target-language examples (few-shot fine-tuning) typically provides a substantial improvement.
How is multilingual QA evaluated?
The standard metrics are Exact Match (EM) — the fraction of predictions that exactly match the gold answer — and token-level F1, which rewards partial matches. Results are typically reported per language and as a macro-average across languages on benchmarks such as MLQA, XQuAD, or TyDi QA.
What is the difference between extractive and generative multilingual QA?
Extractive QA selects a contiguous span from the given passage as the answer; generative QA produces a free-form answer string, which is more flexible but harder to evaluate and more computationally demanding. Most benchmark datasets are designed for extractive evaluation, but generative models like mT5 can handle both styles.
Does multilingual QA work for languages not seen at fine-tuning time?
If those languages are present in the backbone model's pretraining corpus, zero-shot transfer is possible but may be weak. Languages outside the backbone vocabulary receive no meaningful transfer and require dedicated pretraining before fine-tuning.
Sources
- Artetxe, M., Ruder, S., & Yogatama, D. (2020). On the cross-lingual transferability of monolingual representations. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 4623–4637). ACL. DOI: 10.18653/v1/2020.acl-main.421 ↗
- Clark, J. H., Choi, E., Collins, M., Garrette, D., Kwiatkowski, T., Nikolaev, V., & Palomaki, J. (2020). TyDi QA: A benchmark for information-seeking question answering in typologically diverse languages. Transactions of the Association for Computational Linguistics, 8, 454–470. link ↗
How to cite this page
ScholarGate. (2026, June 3). Multilingual Question Answering (Cross-lingual MRC). ScholarGate. https://scholargate.app/en/deep-learning/multilingual-question-answering
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- BERT-based ClassificationDeep learning↔ compare
- Multilingual Sentence EmbeddingsDeep learning↔ compare
- Multilingual TransformerDeep learning↔ compare
- RoBERTa-based ClassificationDeep learning↔ compare