CSDQA: Diagram Question Answering in Computer Science

Shaowei Wang, Lingling Zhang, Yi Yang, Xin Hu, Tao Qin, Bifan Wei, Jun Liu

Abstract: Visual Question Answering (VQA) has been a research focus of the computer vision community for recent years. Most of them are accomplished and verified on images of natural scenes. However, Diagram Question Answering (DQA) which is the task of answering natural language questions based on diagram is rarely noticed. Diagram is a more abstract carrier of knowledge and important resource composition in the multi-modal knowledge graph, research on it is of great significance for understanding the cognitive behavior of learners. In order to fill the scarcity of such data, this paper proposes the Computer Science Diagram Question Answering (CSDQA) dataset, which is the first geometric type of diagram dataset in this field. This dataset contains 1,294 diagrams with rich fine-grained annotations and 3,494 question-answer pairs, including multiple choice and true-or-false questions with two levels of difficulty. We have open sourced all the data in http://zscl.xjtudlc.com:888/CSDQA , hoping to provide convenience for researchers and make it the high-quality data foundation of DQA.