Unit conflict resolution for automatic math word problem solving

Abstract

Among the statistical approaches for math word problem solving, template based approaches have shown to be more robust against a wide spectrum of math word problems, while other approaches target simple arithmetic problems that compose of only one operation or equation. However, even template based systems are poor in performance for questions that contain different units to describe the same measurement. This paper presents a unit conflict resolution system to improve the performance and accuracy of template based systems under minimal supervision. To illustrate the importance of unit conflict resolution for math word problems, we have annotated a new dataset of 385 algebra word problems. We evaluate the performance of our approach both on a benchmark dataset and this new dataset. Experimental results show that integration of our system to an existing automatic math word problem solver outperforms state-of-theart results when the dataset contains different units to describe the same measurement.