AUTOMATIC SOLVING OF PHYSICS WORD PROBLEMS

Abstract

We present a system that solves simple physics word problems (PWPs) stated in the English language. The main feature of the system is that it is deterministic and yields a correct solution based on real physics formulas. The program generates the solution in a tabular form, showing givens, unknowns, and solution steps. We performed a thorough analysis on the previous work in this field. Most of the research was accumulated in math word problem (MWP) solvers. We have found that these programs are not capable of solving problems from Ukrainian physics textbooks. We identified several types of physics problems: theoretical, value conversion, value comparison, unknowns finding, value change. We developed separate problem-solving strategies for each type. The program uses named entity recognition (NER), a technique in natural language processing (NLP), to identify key elements in the problem text. We created a set of rules for marking these entities. Then, problem type recognition is performed. Depending on the type, the list of entities is converted into the internal problem representation. Value conversion and comparison problems are easy to handle. We developed a recursive algorithm for solving unknowns-finding problems which turned out to be a simplified version of Stanford Research Institute Problem Solver (STRIPS) algorithm. However, developing a universal algorithm for solving value-change problems presents a significant challenge. We believe this problem type belongs to the NP-hard class, indicating inherent difficulty in finding optimal solutions. The interface of the program is a web-application. The user can type the problem text and see the solution on a web page. Additionally, the result of NER is presented. Constructing a general problem solver is challenging. While our program can solve basic physics problems, complex problems involving forces, energy, etc., remain unsolved. However, our solver has great potential for future development. We have thoroughly analyzed its capabilities and limitations and proposed ideas for future research.