Question Answer Dataset for Machine Reading Comprehension
Project no. 20756 · Status: Kapanmış (Yayın bekleniyor) · 2019
Summary
Advances in the language modeling field in recent years have progressed the outcome of many natural language processing problems. One of these problems, the “Machine Reading Comprehension (MRC)” has made significant improvement due to large-scale, high-quality datasets. However many studies have been done for English because many datasets are available in English, and studies on other languages such as Turkish are so rare. To address this issues, Turkish Biology question and answer dataset (TurBiQuAD) has been developed in this project. The dataset, consisting of 13153 samples in total, extracted from high school biology textbooks and Wikipedia articles in the field of biology. Since the dataset is in specific-domain, biology teachers whose mother tongue is Turkish are hired in order to prepare higher quality questions and answers. The dataset is analyzed from four different aspects: general statistics, question types, answer types and difficulty level analysis. According to the analysis, the most common question types are “what” and “which” question types. “Why” and “how” question types cover 18% of the data. The “who”, “where”, and “when” question types contain only 6% of them, due to the characteristics of biology texts. To the best of our knowledge, TurBiQuAD is the first short-answer, medium-sized, domain-specific MRC dataset in Turkish.