Bidirectional Training for Generating Math Word Problems Using Pre-Trained Model and Prompt

Math word problems (MWPs) are a vital type of exercise that greatly enhances students' mathematical proficiency and problem-solving abilities. This paper aims to explore the challenging task of automatically generating MWPs, which necessitates the integration of mathematical knowledge, natural language processing, and logical reasoning. We propose an innovative approach that utilizes bidirectional training and prompt construction to fine-tune pre-trained language models. We evaluate our framework on multiple datasets and metrics, demonstrating its effectiveness and superiority over baselines. Our framework introduces a novel tool and perspective for mathematical education and research, creating new possibilities for generating various types of MWPs.