ArQ: Arabic Question Answering Dataset

Summary

ArQ is a question answering dataset in Levantine Spoken Arabic and Modern Standard Arabic (MSA), consisting of 32,625 triplets (context-question-answer).

Introduction

The dataset follows the format and methodology of HeQ (Hebrew Questions and Answers Dataset). A team of annotators were given random context paragraphs in either spoken Arabic or MSA, and were asked to write relevant questions and mark the correct answers. The answer to each question was segment of text (span) included in the relevant paragraph.

Paragraphs were sourced using two types of sources: (1) for MSA we used short news articles from an online Israeli-Arabic weekly newspaper, and (2) for spoken Arabic we used transcriptions of short videos and recorded interviews in Levantine Arabic.

Questions on both sources were written in Levantine Spoken Arabic (no MSA questions were written).

Question Features

Two types of questions were collected:

Answerable questions (24,124; 74%): Questions for which a single correct answer is present in the paragraph.

Unanswerable questions (8,501; 26%): Questions related to the paragraph's content, where a correct answer is not present in the paragraph, but the paragraph provides a "plausible" incorrect answer in terms of logic.

Quality Labels

As part of ongoing quality control during the collection process, and additional checks on the test and validation sets, approximately 12% of the final data was manually chekced for quality.

Triplets received one of the following quality labels:

Verified: Questions that passed the threshold and were relatively easy, with wording exactly or similar to the relevant sentence in the paragraph, or very common questions.

Good: Questions with wording that was significantly different (lexically or syntactically) from the wording of the relevant sentence in the paragraph.

Gold: Questions that involve more complex inference-making.

Rejected: Questions that did not pass the threshold and therefore not indcluded in the published data.

Additional Answers

After splitting the data, the test and validation subsets underwent additional processing. In cases where there were multiple correct answer spans for an answerable question, the additional possible answer spans were added by annotators to both subsets to enhance robustness.

For example, if the answer appears in quotation marks, another possible answer could be the same answer without the quotation marks. Another example involves answers that may or may not include prepositions preceding the content or appositions. Each answerable question in the test and validation sets received 0 to 3 additional possible answers.

Dataset Statistics

The table below shows the number of answerable and unanswerable questions, by source:

	MSA	Spoken	Total
Answerable	12421	11703	24124 (74%)
Unanswerable	4425	4076	8501 (26%)

The table below shows the number of triplets, by sub-set:

	MSA	Spoken	Total
Train	15080	14197	29277 (90%)
Val	928	745	1673 (5%)
Test	838	837	1675 (5%)

The table below shows the number of unique questions and paragraphs, by source:

	MSA	Spoken
Questions	16846 (52%)	15779 (48%)
Unique Paragraphs	1016	1024

The table below shows the question word distribution in the dataset:

What	Who	How Much/Many	Which	Where	When	How	Why
11517 (34%)	7323 (22%)	4451 (13%)	4103 (12%)	3293 (10%)	1351 (4%)	880 (3%)	700 (2%)

Code

upcoming.

Model

upcoming.

Contributors

ArQ was annotated by Webiks for MAFAT, as part of NNLP-IL, the Israeli national initiative in the field of NLP in Hebrew and Arabic.

Contributors: Amir Shufaniya (Webiks), Carinne Cherf (Webiks) and Yossy Eizenrouah (MAFAT).