融合情感的开放域对话翻译系统设计  

Design of an open - domain dialogue translation system integrating emotion

作　　者：王琪[1] WANG Qi(Harbin Medical University,Daqing Heilongjiang 163000,China)

机构地区：[1]哈尔滨医科大学,黑龙江大庆163000

出　　处：《自动化与仪器仪表》2023年第2期226-230,共5页Automation & Instrumentation

摘　　要：针对在人机交互中,因中文特殊性,导致对话翻译不自然、不流畅等问题,提出一种融合情感的开放域对话翻译系统。为尽可能地挖掘样本数据信息,通过词汇级别向量和字符级别向量相结合的方式,解决分词处理中的OOV问题。并利用双向LSTM解码器,联系单词上下文解码文本隐藏向量。结合引入注意力机制的LSTM解码器,对对话情感分布进行标注,并通过情感判断辅助模型,比较输入语句和输出翻译之间的情感距离,帮助模型确定对话的情感分布。测试结果表明,在对翻译效果的自动测评中,无论是单独单词还是双单词的测评上,融合情感的开放域对话翻译系统比Seq2Seq基本序列模型分别高出0.06和0.3个百分点,而和人工翻译相比则有一定差距;在翻译效果的人工测评中,内容全面性上融合情感的开放域对话翻译系统比Seq2Seq基本序列模型高出0.11-0.29个百分点,情感准确性上情感的开放域对话翻译系统比Seq2Seq基本序列模型高0.03-0.16,两方面较人工翻译都还有不足。这说明无论是在内容还是情感上,情感的开放域对话翻译系统都优于同类型其他翻译模型,具有一定的实用价值。但相较于人工翻译,则还有上升空间,需要进一步的研究。In order to solve the problem that dialogue translation is not natural and fluent due to the particularity of Chinese in human-computer interaction, an open field dialogue translation system based on emotion fusion is proposed. In order to mine the sample data information as much as possible, the OOV problem in word segmentation is solved by combining lexical level vector and character level vector. And the bidirectional LSTM decoder is used to decode the text hidden vector in relation to the word context. The LSTM decoder, which introduces attention mechanism, is used to annotate the affective distribution of the dialogue, and the affective distance between the input statement and the output translation is compared by the affective judgment auxiliary model to help the model determine the affective distribution of the dialogue. The test results show that, in the automatic evaluation of the translation effect, the open-domain dialogue translation system with emotion fusion is 0.06 and 0.3 percentage points higher than the Seq2Seq basic sequence model, respectively, but has a certain gap compared with human translation. In the artificial evaluation of translation effect, the open-domain dialogue translation system with emotion integration is 0.11-0.29 percentage higher than the Seq2Seq basic sequence model in terms of comprehensiveness of content, and 0.03-0.16 percentage higher than the Seq2Seq basic sequence model in terms of emotion accuracy, both of which are still insufficient compared with the human translation. This shows that the open field dialogue translation system of emotion is superior to other translation models of the same type, both in terms of content and emotion, and has certain practical value. However, compared with human translation, there is still room for improvement and further research is needed.

关 键 词：对话翻译 分词处理 情感判断 情感相似度 隐含信息 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]