一种规则与统计相结合的应用题句子语义角色识别方法

Hybrid method based on rules and statistics for semantic role annotation in math word problems

以应用题自动求解为目标,以高考入学考试数学试卷中的分层抽样应用题为研究对象,重点研究了分层抽样应用题的句子语义角色识别方法。根据分层抽样的原理,首先定义了分层抽样题意表征中的五种核心语义角色,分别为总体、样本、总体中的层、样本中的层和实体之间的关系。基于这五种语义角色,应用题题意理解中的核心问题被转换为对应用题文本中的句子进行语义角色判定。提出了一种基于特征词与n-gram模型相结合的句子语义角色判定方法,对分层抽样应用题文本中的句子进行语义角色判定。根据测试集中的实验结果,应用题的整题识别准确率由基于特征词的判定方法的17.95%提高到64.1%。实验结果说明基于特征词与ngram模型相结合的句子语义角色判定方法能够提高题意理解的准确率。

This paper proposed a semantic role annotation method of sentences for the stratification sampling word problem in China’s college entrance examination. According to the basic principles of stratification sampling, this paper defined five core semantic roles to represent the meaning of stratification sampling word problem. These five roles were population, levels in the population, sample, levels in the sample and the relation between entities. With the help of these five semantic roles, it could solve the word problem if the sentences in the word problem could be mapped with these semantic roles. To achieve this goal, this paper presented a hybrid method based on characteristic words and n-gram model to annotate the semantic roles of sentences in stratification sampling word problem. The experimental results show that the accuracy of the right-annotated word problems improves from 17.95% to 64.1%. It proves that the hybrid method can significantly improve the accuracy of stratification sampling word problem understanding.