CASIA OpenIR  > 毕业生  > 硕士学位论文
Alternative TitleResearch on the Sentence clustering for Automatic Essay Scoring
Thesis Advisor徐波
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline软件工程
Keyword自动作文评分 聚类 编辑距离算法 语义相似 词性序列相似度 依存结构相似度 树的编辑距离算法 句子流利度 Automatic Essay Scoring Clustering Edit Distance Algorithm Semantic Similarity The Part-of-speech Sequence Similarity The Dependency Structure Similarity The Tree Edit Distance Similarity Sentence Fluency
Abstract大规模的语言类考试中的作文阅卷问题,一直是令人头疼的事,既耗费人力物力,又不能保证很高的准确性;而自动作文评分系统因为其快速、客观、一致性好等特点已引起计算机科学界和教育界越来越多的关注,作为该研究领域的领头人,美国早在上世纪90年代就已经推出了多款商用的自动作文评分系统,其中一些已经应用于我们所熟知GMAT和TOEFL考试中,用来对考生作文进行评分;而日本、韩国、台湾等国家和地区也加大了对该领域的研究投入,并相继推出了一批实用的或实验性的产品。但是目前已推出的自动作文评分系统大多是针对采用母语写作的作文的评分(TOEFL虽针对第二语言作文,但其应试者水平都较高),很少是针对采用第二语言写作的作文的,已有资料表明,针对母语写作的自动作文评分系统在对第二语言作文进行评分时,性能很差。这主要是因为采用母语写作的作文,通常语句都比较通顺,而采用第二语言写作的作文,通常句子的语法和用词上都有很多问题,这些问题会造成评分系统在提取特征时出现较大误差,导致评分不准确。因此,要解决对第二语言作文的自动评分问题,就必须对这类作文的句子进行深入的分析。 针对上述问题,本文的研究工作主要是对初级的汉语学习者的作文进行句子质量方面的分析。考虑到这些初级的汉语学习者所掌握的句子写法的种类有限,本文尝试采用一种模仿词典的方式,通过对大量的句子进行聚类并构造“句典”,然后由专家对句典中的各类句子进行打分以分出每一类句子的优劣。论文的主要工作如下:  对句子按照语义相似性进行聚类,通过两两比较句子的语义相似度,将语义相似的句子归为一类,实现了四种语义相似度计算方法,并比较了不同的语义相似度计算方法对聚类性能的影响。  对句子按照结构相似性进行聚类,通过两两比较句子的结构相似度,将结构相似的句子聚成一类,实现了两种结构相似度的计算方法,并比较了不同的结构相似度计算方法对聚类性能的影响。  对句子按照写作水平进行聚类,这一部分的工作主要集中在对句子流利度特征的提取以及对比采用不同的特征对聚类性能的影响。其中对流利度特征的提取分别采用了基于语言模型的办法和基于搜索引擎的办法,而比较聚类性能时,主要展示了流利度特征的准确性对聚类性能的影响。 实验表明,在语义聚类中,采用编辑距离算法配合人工词典的语义相似度计 算方法优于其他的方法,而本文所采用的基于标杆的聚类算法在某些情况下也好于常用的组平均聚类算法(Groupwise-average);而在结构聚类中,采用词性序列的聚类方法略好于采用依存结构的聚类方法;最后在写作水平聚类中,通过比较不同的流利度特征提取方法对聚类性能的影响,发现流利度特征越准确,聚类性能越好。
Other AbstractScoring essays by human in a large-scale language test is often thought to be a headache. On one hand, it consumes a lot of manpower and material resources, on the other hand human often grade essays subjectively. But the Automatic Essay Scoring (AES) system doesn’t have these problems, it can assess essays rapidly, objectively and consistently, so it has brought more and more attention to the field of computer science and education. As a leader in the AES system research area, the American scientists developed several commercial systems in the 1990th, and some of them are used in the well-known GMAT and TOEFL test to score the essays. In recent years, Japan, Korea and Taiwan etc. also increased investment in this research area, and developed some practical or experimental AES systems. However, as so far, most of the available AES systems are designed to score the essays which are written by native speakers (although TOEFL is a test to assess foreign English learners, but the examinees’ English level is higher than normal people’s), and few of the systems are used to assess essays which are written by non-native learners. Some statistics have showed that while using an AES system which is originally designed to assess native speakers to assess non-native learners, it often gets low performance. This is because when a native speaker writes an essay, he/she can write the sentence fluently, but when a non-native learner writes an essay, he/she always makes some grammatical errors in sentences or uses wrong words, which makes the AES system can’t extract features correctly and score precisely. So, the key to solve the problem of automatically score non-native learners’ essays is to analyze these essays’ sentences in-depth. So this paper mainly studies how to analyze the quality of the sentences which are written by low level Chinese learners. As the learners of this level haven’t mastered large numbers of style of writing yet, it’s possible to construct a sentence-dictionary by clustering similar sentences into several groups, after that, experts can score every group of sentences to distinguish good sentences from bad ones. The main work of this paper is listed below:  Cluster sentences into groups by comparing each pair of sentences’ semantic similarity, and implement four algorithms to calculate the semantic similarity of two sentences. It shows that using different similarity algorithm in the clustering algorithm, the final clustering ...
Other Identifier200828009029084
Document Type学位论文
Recommended Citation
GB/T 7714
赵知. 用于自动作文评分的句子聚类研究[D]. 中国科学院自动化研究所. 中国科学院研究生院,2011.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20082800902908(736KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[赵知]'s Articles
Baidu academic
Similar articles in Baidu academic
[赵知]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[赵知]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.