生成,推理与排序:基于多任务架构的数学文字题生成  

Generating,Reasoning and Ranking:Multitask Learning Framework for Math Word Problem Generation

作　　者：曹天旸 许晓丹 常宝宝[1,3] CAO Tianyang;XU Xiaodan;CHANG Baobao(National Key Laboratory of Multimedia Information Processing,School of Computer Science,Peking University,Beijing 100871,China;School of Software and Microelectronics,Peking University,Beijing 100871,China;Jiangsu Collaborative Innovaton Clnter for Language Ability,Jiangsu Normal University,Xuzhou,Jiangsu 221009,China)

机构地区：[1]北京大学计算机学院多媒体信息处理全国重点实验室,北京100871 [2]北京大学软件与微电子学院,北京100871 [3]江苏师范大学江苏省语言能力协同创新中心,江苏徐州221009

出　　处：《中文信息学报》2024年第12期159-169,180,共12页Journal of Chinese Information Processing

基　　金：国家自然科学基金(61936012,61876004)。

摘　　要：数学文字题是一段能反映数学等式潜在逻辑的叙述性文本。成功的数学问题生成在语言生成和教育领域都具有广阔的应用前景。前人的工作大多需要人工标注的模板或关键词作为输入,且未考虑数学表达式本身的特点。该文提出了一种多任务联合训练的问题文本生成模型。作者设计了三个辅助任务,包括数字间关系抽取、数值排序和片段替换预测。它们与生成目标联合训练,用以监督解码器的学习,增强模型对运算逻辑和问题条件的感知能力。实验证明,该文所提方法能有效提升生成的数学文字题的质量。A math word problem(MWP)is a narrative which reflects the underlying logic of math equations.Successful MWP generation has wide prospect in language generation and educational field.This paper proposes a multitask learning based MWP generation framework.We devise three novel tasks,including number relation extraction,number ranking and sentence substitution prediction.These tasks are jointly trained with generation objective and supervise the learning of MWP decoder,so as to enhance the model’s comprehension of arithmetic logic and condition.Experiments demonstrate the effectiveness of our proposed method in equation consistency of generated MWPs.

关 键 词：数学文字题生成 多任务学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]