Keywords: reasoning, generation, question-answering
Abstract: A wide range of approaches exists for automatically solving math word problems (MWPs), with the majority focusing on obtaining the final correct final answer, which is not enough. Solutions with step-by-step explanations are valuable in many applications, especially in education, to help students better comprehend problem solving strategies. Recent approaches built on large-scale, pre-trained language models, offer a possibility not only to reach a single final answer but also to generate intermediate solution steps, leveraging the language understanding and generation capabilities of language models. However, language models lack mathematical reasoning ability and often cannot generate coherent steps with a clear solution strategy. Thus, we study the problem of fine-grained, step-by-step controllable solution generation for MWPs. We explore whether we can learn a solution strategy that informs future steps and apply controllable generation methods to generate step-by-step not word-by-word. We (i) train a math operation predictor to plan the mathematical operation to apply in the next step given history steps and (ii) use the predicted operation to prompt a language model to generate the next step token-by-token. We conduct numerical experiments on the GSM8K dataset and show that our method improves the overall MWP solving accuracy and solution interpretability with step-by-step plans.
Paper Type: long
Research Area: NLP Applications
0 Replies
Loading