Machine Solver for Physics Word Problems

Megan Leszczynski, Jose Moreira

Abstract:

We build a machine solver for word problems on the physics of a free
falling object under constant acceleration of gravity. Each problem
consists of a formulation part, describing the setting, and a question
part asking for the value of an unknown. Our solver consists of
two long short-term memory recurrent neural networks and a numerical
integrator. The first neural network (the labeler) labels each
word of the problem, identifying the physical parameters and the
question part of the problem. The second neural network (the
classifier) identifies what is being asked in the question. Using
the information extracted by both networks, the numerical integrator
computes the solution. We observe that the classifier is resilient
to errors made by the labeler, which does a better job of identifying
the physics parameters than the question. Training, validation and test
sets of problems are generated from a grammar, with validation and test
problems structurally different from the training problems. The overall
accuracy of the solver on the test cases is 99.8%.

TL;DR: We build an automated solver for a class of physics word problems, using a combination of neural networks and a numerical integrator.

Conflicts: ibm.com, cornell.edu