Solving Math Word Problems by Combining Language Models With Symbolic Solvers

This is the repo for the paper: Solving Math Word Problems by Combining Language Models With Symbolic Solvers.

Usage

To solve a math word problem, we first formalize the word problem as a set of variables and equations and then use SymPy to solve the equations.

Set the OpenAI key before running the script: export OPENAI_API_KEY='sk-...'

from utils import *
from prompts.declarative_three_shot import DECLARATIVE_THREE_SHOT_AND_PRINCIPLES

question = 'xxx'
eq_list = get_declarative_equations(model='code-davinci-002', question=question, prompt=DECLARATIVE_THREE_SHOT_AND_PRINCIPLES, max_tokens=600, stop_token='\n\n\n', temperature=0)
answer = get_final_using_sympy(eq_list)

The get_declarative_equations method calls the OpenAI API to generate the equations, and the get_final_using_sympy method solves the equations to get the final answer (see an example in run.py).