面向汉译英口语测试中自动评测方法的研究

The automatic scoring technology of English oral examination is a technical means to use computer to automatically score candidates' oral responses. The existing research on the automatic score of spoken English test is a relatively mature technology since the questions in the test are mostly the "controlled type", namely reading aloud; therefore candidates’ responses are highly restricted to standard keys. However, research on the automatic score of the "semi-controlled type" questions, such as translation and story- retelling, are relatively insufficient due to the fact that there is no exclusive standard key to the questions. The evaluation of Chinese-English translation in the oral examination is even more problematic since there is no specific evaluation strategy. . .This project intends to combine intelligent signal processing and natural language processing technologies to study the key techniques of automatic scoring for Chinese-English oral translation. The principal evaluation criteria of Chinese-English translation include the key information of the original language as well as the accuracy and fluency of the target language. Therefore, the key problems that we have to solve include key information identification, synonym recognition, and semantic analysis of the sentence and pronunciation fluency determination. .Through the establishment of the standard answer speech database to oral examinations, this project intends to achieve the following goals: 1) to explore the deep learning method to improve the recognition rate of key information; 2) to draw a semantic tree to help improve the processing technology of synonyms and sentence order; 3) to study English speech and grammar based on part of speech and syntax Analysis; 4) to study the selection of fluency indicators and feature quantification method. The project is expected to be able to eventually determine a set of fractional fusion method based on machine learning.

英语口语考试自动评分是利用计算机对考生口语答卷进行打分。现有关于口语自动评分的研究大多是针对朗读题等限定内容的“控制型”题型，技术相对成熟；而针对翻译题、复述题等未给定标准答案的“半控制型”题型的研究成果相对较少，特别是汉译英题型，更是缺乏具体的评测策略。基于此，本项目拟结合智能信号处理和自然语言处理等技术，研究针对汉英口语翻译题的自动评分关键技术。汉译英题型的评分要点主要包括答案中涉及的关键信息点、汉英表达的准确性和合理性以及句子表达的流畅性。所以要解决的关键问题主要有答题关键信息识别、同义词辨析、句子语法分析和发音流利度判定等。本项目将通过建立口语试卷的标准答案语音库，探讨深度学习方法提高关键信息识别率；建立试题答案关键信息语义树，探讨答题同义词处理及句子词序识别技术；研究基于词性标注和句法特征的英语语音语法分析；研究流利度指标选取及特征量化方法；确定一套基于机器学习的分数融合方法。

本项目重点研究口语考试中“半控制型”题型-翻译题的自动评分问题。特别是口语汉译英题型的评分方法的研究。研究中结合智能信号处理和自然语言处理等技术。分析了人工评分中汉英口语翻译题的评分技术，我们认为汉译英题型的评分要点主要包括答案中涉及的关键信息点、汉英表达的准确性和合理性以及句子表达的流畅性。所以评分中要解决的关键问题主要有答题关键信息识别、同义词辨析、句子语法分析和发音流利度判定等。探讨建立口语试卷的标准答案语音库，探讨深度学习方法提高关键信息识别率；建立试题答案关键信息语义树，探讨答题同义词处理及句子词序识别技术；研究基于词性标注和句法特征的英语语音语法分析；研究流利度指标选取及特征量化方法；研究了一套口语汉译英自动评分方法，并确定一套基于机器学习的分数融合方法。整理了在研究过程中构建的超百GB规模的、已标注语料库。本项目研究期内，研究成果申请发明专利7项。登记软件著作权11项。发表重要国际、国内期刊及会议论文17篇，其中SCI/EI 检索论文15篇。培养博士研究生1名，硕士研究生4名。参加有影响的国内外学术会议11次，累计邀请3位国内外专家举办学术交流报告5次，有效地促进了相关方向的研究和发展。