基于回归分析的中文作文自动评分技术研究

发布时间：2018-05-04 21:48

本文选题：自动评分 + 回归分析　；参考：《哈尔滨工业大学》2016年硕士论文

【摘要】：评分在教学过程中有着重要的地位,他是考核学生学习效果的手段,也是考查教师教学水平的必要方式,而评分最关注的是评分的公平性与客观性。随着人工智能的快速发展,利用人工智能的机器自动评分逐渐走入人们的视野,不难看出,机器自动评分技术具有人工评分不具备的客观性、中立性,同时高效而省时省力,可以反复的完成评卷工作等,具有诸多优异的特性。高考是全省乃至全国范围内的大型升学资格考试,每年愈百万人参加,它决定了考生选择大学和进入大学的资格标准,在我国有着不可替代的重要作用。高考中文作文评分为评分的一个特殊应用场景,由于绝大部分考生受到了良好的教育与培养,作文评分的评分对象有着较强的语言能力,语文作文考查考生语言能力的同时还需要考查考生的理解能力、逻辑思维能力、文学素养、知识积累等多各方面内容。然而,与高考评分的严肃性相对的,是高考中文作文评分紧迫的阅卷时间,评卷老师工作量极大,可以说作文人工评卷是对高考作文评分公平性与权威性的极大挑战。本文尝试使用自动评分技术对高考中文作文进行自动评分。为解决人工评分局限性,首先需要理解、分析中文作文考试以及评分过程的复杂性,也是本课题面对的机遇与挑战。本文将着重探讨基于回归分析的中文作文自动评分的核心技术,主要内容包括了:(1)评分过程概念的分离与抽象,提出合理的评价指标与目标优化参数,为后续实验与实验工作的评价建立了基础;(2)采用启发式的特征抽取方法,建立基于浅层语言特征的回归模型,设定了实验的基线(baseline)。归纳出本课题基于启发式特征抽取的基本实验方法与思路;(3)建立基于启发式的深层语言特征的回归模型,对深层的语言特征进行分析、拆分与抽取,并对实验结果进行了合理的归因,更加深入的理解问题本质,了解了作文评分的难点与要点,使得后续的工作更加聚焦。(4)尝试使用深度学习的方法,用篇章向量来表示文章,利用回归模型进行训练和测试,以探索自动特征抽取方法对系统的影响,从而突破系统某些的特定指标的瓶颈。
[Abstract]:Scoring plays an important role in the teaching process. It is not only a means to assess the students' learning effect, but also a necessary way to check the teachers' teaching level, and the most concern of the scoring is the fairness and objectivity of the scoring. With the rapid development of artificial intelligence, the machine automatic scoring using artificial intelligence has gradually come into people's view. It is not difficult to see that automatic machine scoring technology has objectivity, neutrality, high efficiency and time saving. You can repeatedly finish marking work, and so on, with many excellent characteristics. The college entrance examination is a large-scale examination of qualification for higher education in the whole province and even all over the country. Every year, more than one million people take part in it, which determines the qualification criteria for candidates to choose universities and enter universities, and plays an irreplaceable role in our country. The Chinese composition score of the college entrance examination is a special application scene of scoring. Because most of the examinees have received good education and training, the scoring object of the composition scoring has a strong language ability. It is necessary to examine the students' understanding ability, logical thinking ability, literary accomplishment, knowledge accumulation and so on while examining the language ability of the examinee in Chinese composition examination. However, in contrast to the seriousness of the college entrance examination score, it is the urgent time to mark the Chinese composition score of the college entrance examination, and the workload of the examination paper evaluation teacher is great. It can be said that the composition manual evaluation is a great challenge to the fairness and authority of the composition score of the college entrance examination. This paper attempts to use automatic scoring technology to automatically score Chinese composition in college entrance examination. In order to solve the limitation of artificial scoring, it is necessary to understand and analyze the complexity of Chinese composition examination and scoring process, which is also the opportunity and challenge faced by this subject. This paper will focus on the core technology of automatic scoring of Chinese composition based on regression analysis. The main contents include the separation and abstraction of the concept of the score process, and the reasonable evaluation index and objective optimization parameters are put forward. A heuristic feature extraction method is used to establish a regression model based on shallow language features and the baseline of the experiment is set up. The basic experimental methods and ideas of heuristic feature extraction are summarized. (3) the regression model of deep language feature based on heuristic is established, and the deep language feature is analyzed, split and extracted. The results of the experiment are attributed reasonably to understand the nature of the problem, understand the difficulties and key points of the composition scoring, make the follow-up work more focused. 4) try to use the method of deep learning and use the text vector to express the article. The regression model is used for training and testing to explore the influence of automatic feature extraction on the system, thus breaking the bottleneck of some specific indicators of the system.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：G633.34;O212.1

【相似文献】