基于表征重述的机器文本理解

Machine Text Understanding Using Representational Redescription Process

传统机器翻译系统缺乏联系上下文形成认知的能力,仅根据对应单词的默认含义进行翻译,容易导致语义错误等问题。通过模拟人的表征重述认知过程,提出一种新的机器文本理解与翻译方法。该方法可通过较少的实例对文本进行理解和翻译,避免出现语义理解错误问题,且无需进行繁杂的语法标注。实验表明,该方法可通过引入习得的常识,使翻译出现歧义错误的概率降低到1%以下,并可标注出不符合常理而又无法找到更好解释的句子。

Traditional machine translation methods often cause semantic errors by simple substitution.Because of lacking common sense,machines merely find corresponding words in target language and substitute them with default meanings.A novel machine text understanding and translation method is advanced by simulating human’s representational redescription process.This method can understand and translate text correctly with fewer instances.Semantic errors can be avoided in this method.Complex syntactic label can be avoided too.According to the experiment result,this method can introduce common sense into the understanding process and get a semantic ambiguity error rate lower than 1%.Furthermore ridiculous translations can be labeled if no better translation can be found.