融入句法结构和摘要信息的文本蕴含识别模型

来源：计算机技术与发展杂志2023年第10期北京时间：

作者：钟茂生;孙　磊;罗贤增;王明文

单位：江西师范大学 计算机信息工程学院,江西 南昌 330200 Author(s): ZHONG Mao-sheng;SUN Lei;LUO Xian-zeng;WANG Ming-wen School of Computer Information Engineering,Jiangxi Normal University,Nanchang 330200,China 关键词: 文本蕴含识别;摘要信息抽取;句法结构;互注意力;自注意力 Keywords: recognizing textual entailment;abstract information extraction;syntactic structure;mutual attention;self-attention 分类号: TP391 DOI: 10. 3969 / j. issn. 1673-629X. 2023. 10. 019 摘要: 文本蕴含识别旨在推断两个句子之间的语义关系,通常分为蕴含、矛盾和中立三种类别。 目前,大多数文本蕴含识别方法都是通过互注意力的方法,判定句子之间的语义关系,这种方法只能捕捉局部交互信息,弱化了全局交互信息。据此,提出了融入句法结构和摘要信息的文本蕴含识别模型,通过结合自注意力和互注意力机制的方式,从句子的全局和局部交互信息考虑,并融入句子的句法结构信息,从而更准确地推测句子之间的语义关系;收集和整理了公务员试题的部分选择题,之后,将该模型和文本蕴含识别的思想应用于这部分试题,在试题中,使用摘要信息抽取的方法,解决公务员试题中题目冗长和答案简短导致的长度不对称问题。 实验结果表明,该模型在公共数据集和公务员试题上的表现,超越了多个基准模型。 Abstract: Recognizing textual entailment aims to infer the semantic relationship between two sentences, usually divided into threecategories of entailment, contradiction and neutral. At present, most recognizing textual entailment methods use mutual attention todetermine the semantic relationship between sentences. This method can?only capture local interaction information and weaken the globalinteraction information. Accordingly,a recognizing textual entailment model that integrates syntactic structure and abstract information isproposed. By combining self-attention and mutual attention mechanisms,the global and local interactive information of sentences?is considered,and the syntactic structure information of sentences is integrated, so as to accurately infer the semantic relationship betweensentences. We collect and sort out some multiple-choice questions of civil servant test questions,and then apply the model and the ideaof textual entailment recognition to these test questions.?In the test questions,the method of abstract information extraction is used tosolve asymmetric length problems caused by long questions and short answers in civil service examination questions. Experimental resultsshow that the proposed model outperforms several benchmark models on public datasets and civil servant test questions.

摘要：文本蕴含识别旨在推断两个句子之间的语义关系,通常分为蕴含、矛盾和中立三种类别。 目前,大多数文本蕴含识别方法都是通过互注意力的方法,判定句子之间的语义关系,这种方法只能捕捉局部交互信息,弱化了全局交互信息。据此,提出了融入句法结构和摘要信息的文本蕴含识别模型,通过结合自注意力和互注意力机制的方式,从句子的全局和局部交互信息考虑,并融入句子的句法结构信息,从而更准确地推测句子之间的语义关系;收集和整理了公务员试题的部分选择题,之后,将该模型和文本蕴含识别的思想应用于这部分试题,在试题中,使用摘要信息抽取的方法,解决公务员试题中题目冗长和答案简短导致的长度不对称问题。 实验结果表明,该模型在公共数据集和公务员试题上的表现,超越了多个基准模型。

关键词：文本蕴含识别;摘要信息抽取;句法结构;互注意力;自注意力