Datasets:

nguyenlab
/

vimqa

language:
  - vi
license: cc-by-nc-sa-4.0
size_categories:
  - 10K<n<100K
task_categories:
  - question-answering
task_ids:
  - extractive-qa
pretty_name: VIMQA
tags:
  - multi-hop
  - vietnamese
  - explainable-qa
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
  - config_name: gold_only
    data_files:
      - split: validation
        path: gold_only/validation-*
      - split: test
        path: gold_only/test-*

VIMQA

VIMQA is a Vietnamese dataset for advanced reasoning and explainable multi-hop question answering. Each question requires combining facts from two different Vietnamese Wikipedia articles, and every example ships with sentence-level supporting facts so a model's reasoning chain can be evaluated, not just its final answer.

The schema follows the HotpotQA convention, so tooling written for HotpotQA transfers with minimal changes.

Usage

from datasets import load_dataset

# Full distractor setting: 10 paragraphs per question, 2 of them gold.
ds = load_dataset("nguyenlab/vimqa")

# Gold-only setting: just the supporting paragraphs.
gold = load_dataset("nguyenlab/vimqa", "gold_only")

Configs and splits

Config	Split	Rows	Paragraphs per question
`default`	`train`	8,041	10
`default`	`validation`	1,003	10
`default`	`test`	1,003	10
`gold_only`	`validation`	1,003	1–2
`gold_only`	`test`	1,003	1–2

The default config is the distractor setting: each question comes with 10 candidate paragraphs, of which only the supporting ones are relevant. The gold_only config contains the same questions with distractors removed, which is useful for isolating reading-comprehension ability from retrieval.

Fields

Field	Type	Description
`id`	`string`	Unique example identifier
`question`	`string`	The Vietnamese question
`answer`	`string`	The answer span, or a yes/no answer (`đúng` / `không`)
`type`	`string`	Reasoning type of the question
`context.title`	`list[string]`	Titles of the candidate paragraphs
`context.sentences`	`list[list[string]]`	Each paragraph, split into sentences
`supporting_facts.title`	`list[string]`	Titles of paragraphs containing supporting facts
`supporting_facts.sent_id`	`list[int32]`	Index into that paragraph's `sentences` list

A supporting fact is the pair (title, sent_id): it points at one specific sentence inside one specific context paragraph.

Example

{
  "id": "aebce1bf-35a3-4e0c-85c1-e59b24dfb48b",
  "question": "Diego Maradona nhỏ tuổi hơn Rutherford B. Hayes phải không?",
  "answer": "đúng",
  "type": "bridge",
  "context": {
    "title": ["PH", "Diego Maradona", "Rutherford B. Hayes", ...],
    "sentences": [["Các dung dịch nước có giá trị pH nhỏ hơn 7 ..."], [...], [...]]
  },
  "supporting_facts": {
    "title": ["Diego Maradona", "Rutherford B. Hayes"],
    "sent_id": [0, 0]
  }
}

To recover the text of the supporting sentences:

def supporting_sentences(example):
    lookup = dict(zip(example["context"]["title"], example["context"]["sentences"]))
    return [
        lookup[title][sent_id]
        for title, sent_id in zip(
            example["supporting_facts"]["title"],
            example["supporting_facts"]["sent_id"],
        )
    ]

Source data

Contexts are drawn from Vietnamese Wikipedia. Questions and supporting-fact annotations were written by human annotators.

Citation

@inproceedings{le-etal-2022-vimqa,
    title = "{VIMQA}: A {V}ietnamese Dataset for Advanced Reasoning and Explainable Multi-hop Question Answering",
    author = "Le, Khang and Nguyen, Hien and Le Thanh, Tung and Nguyen, Minh",
    booktitle = "Proceedings of the Thirteenth Language Resources and Evaluation Conference",
    month = jun,
    year = "2022",
    address = "Marseille, France",
    publisher = "European Language Resources Association",
    url = "https://aclanthology.org/2022.lrec-1.700",
    pages = "6521--6529",
}