Solving Math Word Problems concerning Systems of Equations with GPT-3

Researchers have been interested in developing AI tools to help students learn various mathematical subjects. One challenging set of tasks for school is learning solve math word problems. We explore how recent advances natural language processing, specifically the rise powerful transformer based models, can be applied learners with such Concretely, we evaluate use GPT-3, a 1.75B parameter model recently released by OpenAI, three related challenges pertaining problems corresponding systems two linear equations. The are classifying problems, extracting equations from and generating For first challenge, define problem classes find that GPT-3 has generally very high accuracy (80%-100%), all but one these classes. second improves number examples provided model, ranging an 31% zero-shot about 69% using 3-shot learning, which further improved value 80% fine-tuning. third able generate 33% 93%, depending on type.