阅读理解中多项选择的三重匹配与对比正则算法

Triple Matching and Contrast Regularization Algorithm for Multiple Choices in Reading Comprehension

多项选择由文章、问题和选项三部分组成,主要任务是根据文章和问题在多个选项中找出正确选项。已有一些算法对文章、问题和选项之间的匹配策略进行了一定程度的研究,但一般使用成对处理或双向匹配方法,无法充分融合文章、问题和选项。为此,提出一种三重匹配策略(TM),使用对比正则(CR)方法对答案进行区分处理,将文章、问题和答案三者中任意一个元素与其它元素进行匹配,以吸收其他两者的语义信息。实验表明,通过CR捕捉并强化正确答案和错误答案之间的差异,能让模型更好地识别正确与错误答案,以期为该领域的研究人员提供参考与借鉴。

Multiple choice consists of three parts:article,question,and option.The main task is to find the correct option among multiple options based on the article and question.Some algorithms have conducted some research on matching strategies between articles,questions,and options,but generally use paired processing or bidirectional matching methods,which cannot fully integrate articles,questions,and options.To this end,a triple matching strategy(TM)is proposed,which uses contrastive regularization(CR)method to distinguish answers and matches any element of the article,question,and answer with other elements to absorb semantic information from the other two.Experiments have shown that capturing and strengthening the differences between correct and incorrect answers through CR can enable the model to better identify correct and incorrect answers,in order to provide reference and inspiration for researchers in this field.