Nitesh Methani*, Pritha Ganguly*, Mitesh M. Khapra, Pratyush Kumar
Department of Computer Science and Engineering,
Robert Bosch Centre for Data Science and AI
Indian Institute of Technology, Madras
* The first two authors contributed equally
Q: What is the total tuberculosis detection rate in Indonesia?
A: 101
Q: In which year was the tuberculosis case detection rate of Somalia minimum?
A: 2000
Q: Across all years, what is the maximum tuberculosis case detection rate of Benin?
A: 55.46
Problem Statement
Our Contributions
224,377 scientific plots on data soured from real world
28.9 million questions based on templates sourced from manually curated crowdsourced questions
questions which have answers from an Open Vocabulary
perception and QA modules for questions that have answers from an Open Vocabulary
achieves best performance on both PlotQA and DVQA dataset
FigureQA vs DVQA vs PlotQA
Q: Is Light Green the minimum?
A: 1.0
Q: What is the value of mad in drop?
A: 7
Q: What is the average number of Hispanic students in school?
A: 51.67
FigureQA
DVQA
PlotQA
FigureQA vs DVQA vs PlotQA
FigureQA
DVQA
PlotQA
Dataset Creation
Collection
& Curation
Plot
Generation
Question
Collection
Templatization & Instantiation
Multistage Pipeline
VED
OCR
SIE
QA
Proposed Model
Question classifier:
deciding whether the question can be answered from a fixed vocabulary or needs more complex reasoning
QA-as-classification:
answers question from a fixed vocabulary
Multi-staged model:
a pipeline of perception and QA modules for answering complex questions.
Results & Analysis
References
[1] K. Kafle, S. Cohen, B. L. Price, and C. Kanan. DVQA: understanding data visualizations via question answering. CoRR, abs/1801.08163, 2018.
[2] P. Pasupat and P. Liang. Compositional semantic parsing on semi-structured tables. In ACL, 2015
[3] S. E. Kahou, A. Atkinson, V. Michalski, ́A. K ́ad ́ar, A. Trischler, and Y. Bengio. Figureqa: An annotated figure dataset for visual reasoning. CoRR, abs/1710.07300, 2017.
[4] S. Ren, K. He, R. B. Girshick, and J. Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 91–99, 2015