Extractive Question Answering for Kazakh Language

Magzhan Shymbayev, Yermek Alimzhanov

2023

Closed access Metadata (CSL-JSON)
DOI 10.1109/sist58284.2023.10223508
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

This article provides research and development of an extractive question answering system based on the BERT-like model for the Kazakh language. Developing an extractive question answering system requires large training datasets - tens of thousands of annotated question-answer pairs. Such datasets are not available in the majority of languages, including Kazakh. To address this issue, the Kazakh Question Answering Dataset (KazQA) is introduced, which is based on the Stanford Question Answering Dataset (SQuAD) and generated through machine translation using the Google Cloud Translation API. Different large pretrained contextual language models are used as the baseline models - ALBERT and multilingual BERT and are compared with the newly trained monolingual Kazakh model KazBERT. The results demonstrate that the proposed approach can effectively generate question answering systems in low-resourced Kazakh language.