Dataset Card for "pquad"

PQuAD Description

THIS IS A NON-OFFICIAL VERSION OF THE DATASET UPLOADED TO HUGGINGFACE BY Gholamreza Dar

The original repository for the dataset is https://github.com/AUT-NLP/PQuAD

PQuAD is a crowd- sourced reading comprehension dataset on Persian Language. It includes 80,000 questions along with their answers, with 25% of the questions being unanswerable. As a reading comprehension dataset, it requires a system to read a passage and then answer the given questions from the passage. PQuAD's questions are based on Persian Wikipedia articles and cover a wide variety of subjects. Articles used for question generation are quality checked and include few number of non-Persian words.

Dataset Splits

The dataset is divided into three categories including train, validation, and test sets and the statistics of these sets are as follows:

+----------------------------+-------+------------+------+-------+
|                            | Train | Validation | Test | Total |
+----------------------------+-------+------------+------+-------+
|      Total Questions       | 63994 |    7976    | 8002 | 79972 |
|   Unanswerable Questions   | 15721 |    1981    | 1914 | 19616 |
| Mean # of paragraph tokens |  125  |    121     | 124  |  125  |
| Mean # of question tokens  |   10  |     11     |  11  |   10  |
|  Mean # of answer tokens   |   5   |     6      |  5   |   5   |
+----------------------------+-------+------------+------+-------+

Workers were encouraged to use paraphrased sentences in their questions and avoid choosing the answers comprising non-Persian words. Another group of crowdworkers validated the questions and answers in the test and validation set to ensure their quality. They also provided additional answers to the questions in test and validation sets if possible. This helps to consider all possible types of answers and have a better evaluation of models.

PQuAD is stored in the JSON format and consists of passages where each passage is linked to a set of questions. Answer(s) of the questions is specified with answer's span (start and end point of answer in paragraph). Also, the unanswerable questions are marked as unanswerable.

Results

The estimated human performance on the test set is 88.3% for F1 and 80.3% for EM. We have evaluated PQuAD using two pre-trained transformer-based language models, namely ParsBERT (Farahani et al., 2021) and XLM-RoBERTa (Conneau et al., 2020), as well as BiDAF (Levy et al., 2017) which is an attention-based model proposed for MRC.

+-------------+------+------+-----------+-----------+-------------+
|    Model    |  EM  |  F1  | HasAns_EM | HasAns_F1 | NoAns_EM/F1 |
+-------------+------+------+-----------+-----------+-------------+
|     BNA     | 54.4 | 71.4 |    43.9   |    66.4   |     87.6    |
|   ParsBERT  | 68.1 | 82.0 |    61.5   |    79.8   |     89.0    |
| XLM-RoBERTa | 74.8 | 87.6 |    69.1   |    86.0   |     92.7    |
|    Human    | 80.3 | 88.3 |    74.9   |    85.6   |     96.8    |
+-------------+------+------+-----------+-----------+-------------+

LICENSE

PQuAD is developed by Mabna Intelligent Computing at Amirkabir Science and Technology Park with collaboration of the NLP lab of the Amirkabir University of Technology and is supported by the Vice Presidency for Scientific and Technology. By releasing this dataset, we aim to ease research on Persian reading comprehension and the development of Persian question answering systems.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Dataset Card for "pquad"