Automatic Hybrid Mathematical Word Problem Solver

Abstract

Math Word Problems (MWPs) are mathematical problems in which significant information is presented in natural language rather than in mathematical notations. Word problems often narrative of some sort, they are also referred to as story problems and may vary in the amount of language used. The proposed system is built as a combination of a verb classification model, an encoder-decoder model, and a regular expression model that takes math word problem as input and generates the corresponding math expression as output. The proposed system also has the ability to categorize whether the given input is an interrogative math word problem or not, it helps to avoid unnecessary non-word problems to the hybrid system. Verb classification model used here is trying to solve some kind of word problems which are in proper syntactic format and the encoder-decoder model can solve all kind of word problems with the help of recurrent neural network. A regular expression model also takes part in the model to identify and solve outlier mathematical expressions. This proposed system only tries to solve math word problems that contain fundamental operators. It provides a way to handle multiple concepts in the same problem while, at the same time, supporting the interpretability of the answer expression. The proposed hybrid system outperforms previous state-of-the-art methods.