自动作文评分模型及方法研究

发布时间：2019-01-16 09:30

【摘要】：写作作为语言考试的一种必备题型,通常需要大规模的作文阅卷,采用人工评分的方法一方面浪费大量的人力、物力,另一方面,人工评分依赖阅卷老师的主观判断,存在一定误差。随着自然语言处理技术的发展,其在句法分析,语义分析、情感分析等方面取得了突破性的进展,利用自然语言处理技术对作文进行自动评分具有重要的意义。传统的自动作文评分方法主要对整篇文章进行统计,提取词法、句法、语义等特征,代入机器学习模型中进行训练,从而预测评分,这种方法通常依赖于人工提取的规则特征,并且只简单的考虑篇章整体质量,没有很好的考虑到其中更细节的信息,比如篇章的上下文信息等。基于这些问题,本文在传统评分方法上,进一步抽取句子层面的特征,并利用篇章的上下文的时序信息,挖掘作文潜在的逻辑性、连贯性对评分模型的影响。另外,针对语言考试中规定题目的写作情景,从考察作文是否符合题意角度出发,研究了作文主题相关度在评分模型中的效果。本文的研究内容主要包括以下三个方面:(1)研究了句表示方法和深度学习模型在自动作文评分中的应用。在写作这个特定情境下,作文更适合看作具备逻辑性的语句序列,为了挖掘句子之间隐藏的逻辑信息,本文主要使用无监督的方法表示句向量,其中包括Doc2Vec技术、递归自编码器等。然后使用深度学习方法构建模型,本文主要基于CNN模型在提取特征上的优势和LSTM模型适合时序性问题的特点,设计并实现了多种模型结构,找到适合该问题的建模方法。并与传统评分模型进行融合,使评分模型效果得到提升。(2)研究了说明文体裁作文的主题相关度在评分问题中的作用。主要通过提取题目文本与作文文本之间的相似度特征,并引入基于主题的语义离散度的概念,分析其对最终评分结果的影响。(3)设计并实现了一个作文自动评分系统。该系统能够对作文质量给出及时的评分反馈,并且在词法、句法、逻辑、主题角度给出相关的反馈。并增加了纠错模块,可以一定程度上提供给用户更好的写作反馈体验。
[Abstract]:As a necessary type of language test, writing usually requires a large scale composition marking. On the one hand, manual scoring method wastes a lot of manpower and material resources, on the other hand, manual scoring depends on the subjective judgment of the marking teacher. There is a certain error. With the development of natural language processing technology, it has made a breakthrough in syntactic analysis, semantic analysis, emotional analysis and so on. It is of great significance to use natural language processing technology to score composition automatically. The traditional automatic composition scoring method mainly carries on the statistics to the entire article, extracts the lexical, syntactic, semantic and so on characteristic, and substitutes in the machine learning model to train, thus predicts the score, this kind of method usually depends on the manual extraction rule characteristic. Furthermore, the overall quality of the text is simply considered, and the more detailed information, such as the context of the text, is not well considered. Based on these problems, this paper extracts the features of sentence level in the traditional scoring method, and uses the temporal information of the context of the text to explore the potential logic and coherence of the composition to influence the scoring model. In addition, according to the writing situation of the prescribed topic in the language test, the effect of the relevance of the composition theme in the scoring model is studied from the point of view of whether the composition conforms to the meaning of the question. The main contents of this paper are as follows: (1) the application of sentence representation and depth learning model in automatic composition scoring is studied. In the particular situation of writing, composition is more suitable to be regarded as a logical sentence sequence. In order to mine the hidden logic information between sentences, this paper mainly uses an unsupervised method to express sentence vectors, including Doc2Vec technology. Recursive self-encoder, etc. Based on the advantages of CNN model in extracting features and the characteristics of LSTM model suitable for temporal problem, this paper designs and implements a variety of model structures to find a modeling method suitable for this problem. And fusion with the traditional scoring model to improve the effectiveness of the scoring model. (2) this paper studies the role of theme relevance in the scoring problem of genre writing. By extracting the similarity feature between title text and composition text, and introducing the concept of topic based semantic dispersion, this paper analyzes its influence on the final score result. (3) A composition automatic scoring system is designed and implemented. The system can give timely scoring feedback on composition quality, and give related feedback in lexical, syntactic, logical and thematic perspectives. An error correction module is added to provide users with a better writing feedback experience to a certain extent.
【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】：2017
【分类号】：TP391.1

【相似文献】