Semi-Supervised Question Answering with Question-Answer Pair Generation in Bengali

Abstract

Although deep learning architectures and large scale datasets have led to great performance on question answering tasks in high resource languages like English, their performance on lower resource languages, like Bengali, is considerably poorer. This is due to the scarcity of labeled data, which can be attributed to the massive amount of human effort and time required to create such datasets. We work towards a translated Stanford Question Answering Dataset (SQuAD) 1.1 in Bengali and ensure that it is of high quality by using a state-of-the-art translation model and a novel embedding based matching approach to properly align the answer spans in the target language (Bengali) in correspondence with the source language, English. We also introduce an end-to-end question answer generation (QAG) system in the Bengali language to generate question answering (QA) datasets for QA models using roundtrip consistency incorporated in a sequence-to-sequence generation task using Googles mT5 model. Additionally, we train 3 different QA models on our Bengali translated dataset achieving EM and F1 scores of 46.1 and 66.2 respectively. Finally, we demonstrate the effectiveness of our QAG model on a sample dataset of news articles in generating domain-specific QA datasets.