【摘要】 本研究基于现有译文自动评分系统中的语义变量，结合神经网络算法的Word2vec模型和Doc2vec模型，综合归纳为N-gram、语义点和文本相似度三类语义变量，共计14个语义变量，并将其纳入说明文、记叙文、议论文三种文体的翻译质量评价体系，构建回归模型。研究结果显示：（1）三类语义变量对译文成绩具有一定的预测性，每个变量与译文成绩的相关系数高低不等；（2）三类语义变量对译文成绩均有一定程度的解释力，但每类变量的单独模型回归效果弱于将所有变量统一纳入的整体模型；（3）将三类语义变量综合纳入的三种文体翻译质量评价模型拟合程度较高，效果显著。语义点与Word2vec文本相似度的贡献度较高，在三个模型中均有体现，N-gram变量受到源文本文体特征的影响，具体表现形式不一。更多还原

【Abstract】 This study combines semantic variables in existing automatic translation scoring systems with Word2vec and Doc2vec neural network models,and comprehensively compares three types of semantic variables, namely N-gram, semantic points,and text similarity, which includes 14 semantic variables related to the three types of variables in the translation quality evaluation system of explanatory,narrative, and argumentative texts, and constructs a regression model. The results show that:(1) The three types of semantic variables have a certain predictive ability for translation scores, but the correlation coefficients between each variable and translation scores vary;(2) The three types of semantic variables have a certain explanatory power for translation scores, but the regression effect of each individual model is weaker than that of the overall model that includes all variables;(3) The translation quality evaluation models for the three types of text that comprehensively include the three types of semantic variables have a high fitting degree and significant effects. Semantic points and text similarity based on Word2vec contribute more, which is reflected in all three models, while N-gram variables are affected by the characteristics of the source text and have various specific manifestations.更多还原