KQuAD

KQuAD is a Khmer Question Answering dataset designed for evaluating and developing machine reading comprehension systems for the Khmer language.

The dataset follows the extractive question answering formulation: given a question and a corresponding context passage, a model must either identify the answer span within the context or determine that the question cannot be answered from the provided context.

KQuAD includes both answerable and unanswerable questions, following a SQuAD v2-style evaluation setting.

Dataset Summary

KQuAD was developed to provide a benchmark for Khmer machine reading comprehension and question answering.

Each example consists of a context passage and a question associated with that context. For answerable questions, the dataset provides the text span containing the correct answer together with its character-level starting position. For unanswerable questions, no answer span is provided.

The dataset can therefore be used to evaluate two related capabilities:

1. Answer extraction — locating the correct answer span from a given Khmer context.
2. Answerability detection — determining whether sufficient information exists in the context to answer the question.

Dataset Structure

Data Instances

Each instance follows a structure similar to:

## For an answerable example:
{
    "id": "example-id",
    "context": "Khmer context passage...",
    "question": "Khmer question...",
    "answers": {
        "text": ["answer text"],
        "answer_start": [123]
    },
    "is_impossible": False
}

## For an unanswerable example:
{
    "id": "example-id",
    "context": "Khmer context passage...",
    "question": "Khmer question...",
    "answers": {
        "text": [],
        "answer_start": []
    },
    "is_impossible": True
}

Data Fields

Field	Description
id	Unique identifier for each question-answer instance
context	Khmer passage containing information relevant to the question
question	Question associated with the context
answers	Answer annotation containing text and answer_start
answers.text	Extractive answer span; empty for unanswerable questions
answers.answer_start	Character offset of the answer in the context
is_impossible	Indicates whether the question is unanswerable from the given context

Dataset Configurations

KQuAD is released in four configurations. All configurations contain the same underlying articles, contexts, questions, answer annotations, and dataset splits. They differ only in the text segmentation format applied to the Khmer text.

Configuration	Description
default	Original Khmer text without additional text segmentation.
com_seg	Khmer text segmented using the compound word segmentation format.
mor_seg	Khmer text segmented using the morpheme word segmentation format.
kcc_seg	Khmer text segmented using the Khmer Character Cluster (KCC) segmentation format.

These configurations allow researchers to use or compare different Khmer text segmentation schemes while keeping the underlying dataset content and train-validation-test splits unchanged.

Dataset Splits

The dataset contains three predefined splits:

Split	Articles	Paragraphs	Questions	Answerable	Unanswerable
Train	96	954	14,621	8,798 (60.17%)	5,823
Validation	28	260	3,752	2,326 (61.99%)	1,426
Test	34	269	3,700	2,300 (62.16%)	1,400
Total	158	1,483	22,073	13,424 (60.82%)	8,649

Loading the Dataset

The dataset can be loaded using the Hugging Face datasets library:

from datasets import load_dataset

# load default configuration
dataset = load_dataset("sopagnaheang/KQuAD")

# load specific configuration
dataset = load_dataset(
    "sopagnaheang/KQuAD",
    name="com_seg"
)

Evaluation

KQuAD follows the standard SQuAD v2-style extractive question answering evaluation.

The primary evaluation metrics are:

• Exact Match (EM): measures the percentage of predictions that exactly match a reference answer after normalization.
• F1 Score: measures token-level overlap between the predicted answer and the reference answer.
• HasAns EM / F1: evaluates performance only on answerable questions.
• NoAns EM / F1: evaluates whether the model correctly predicts that no answer is available for unanswerable questions.

Baseline Results

As an initial benchmark, four pretrained multilingual encoder models were fine-tuned and evaluated on KQuAD under the same extractive question answering setting: LaBSE, XLM-RoBERTa-base, mmBERT-base, and BGE-M3.

Model	EM	F1	HasAns EM	HasAns F1	NoAns EM	NoAns F1
LaBSE	67.46	76.65	54.13	68.92	89.36	89.36
XLM-RoBERTa-base	64.97	74.20	51.04	65.88	87.86	87.86
mmBERT-base	49.95	54.70	24.13	31.78	92.36	92.36
BGE-M3	67.30	76.38	53.74	68.35	89.57	89.57

Among the evaluated models, LaBSE achieved the best overall F1 score of 76.65, while BGE-M3 performed very closely with 76.38 F1. LaBSE also achieved the strongest performance on answerable questions, with 54.13 HasAns EM and 68.92 HasAns F1.

The baseline results also show a clear difference between answerable and unanswerable questions. All evaluated models perform considerably better on NoAns examples than on HasAns examples, indicating that detecting the absence of an answer is generally easier than locating and extracting the correct answer span from Khmer context passages.

Intended Uses

KQuAD is intended primarily for research on:

• Khmer extractive question answering
• Khmer machine reading comprehension
• low-resource natural language processing
• cross-lingual transfer learning
• evaluation of pretrained multilingual language models
• answerability and unanswerable-question detection

The dataset may also be useful for comparing the Khmer capabilities of multilingual pretrained language models.

Licensing

KQuAD is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

The dataset includes content derived from publicly available Khmer-language sources. Original source materials may remain subject to their respective copyright, licensing, and attribution requirements.

Citation

If you use KQuAD in your research, please cite:

@article{TODO,
  title   = {[TODO: Paper title]},
  author  = {[TODO: Authors]},
  journal = {[TODO]},
  year    = {[TODO]}
}