Reseach On Automatic Scoring Algorithm Based On Semantic Dependency Tree

With the popularity of the test automation, computer scoring has been gradually applied to daily online exam. In general, the type of exam questions are divided into two kinds of objective questions and subjective questions. Currently, automatic scoring of objective questions of computer technology is very mature and widely used in practice. The short answer, essay questions, subjective questions, since the answer is a complete natural language description of the ratings of these questions may relate to natural language processing, pattern recognition, artificial intelligence or even knowledge of, and therefore to implement very difficult. How to effectively correct score of subjective questions automatically, either in practice or in academic fields, is a serious problem.Based on the study of subjective questions at home and abroad automatic scoring based on relevant research, summed up the current commonly used method is to calculate the reference text similarity between answers and candidates answer. In the text similarity computing, most of the research is carried out based on keyword matching. The keyword matching generally have three directions. One is through speech, syntactic dependency analysis or statistical aspects of technology for keyword text screening, the final selection of the most representative of the text semantic set of keywords; the other is a synonym for the keyword expansion thereby increasing the key matching range of words; Third, the sentence is to shut some abstract pattern composed by a keyword, and then in this mode as the basic unit match. The above methods are trying to find the most suitable method for characterizing semantics, but due to the complexity of natural language, the effect of these methods are not ideal.This paper summarizes the existing research results, based on the proposed use of semantic dependency analysis to construct the semantic dependency tree, and thus semantic dependency tree to characterize a sentence semantic approach. And the existing text representation is different, characteristic of this method lies in the ability to characterize backbone of the sentence semantic information, and original text expression that is the backbone of the sentence semantics can be derived mutual. On this basis, the paper proposes a further representation of text in a text-based semantic dependency tree semantic similarity calculation method. This method is divided into words similarity level calculation Sentence similarity computing and full-level similarity calculation in three stages. Among them, the similarity of the words used to calculate the level of existing research results. In the sentence similarity computing, text in hand scoring combined experience in the premise, explained thedifference between text similarity and the similarity of answers for problems existing text simply as a score based on the similarity of the methods would provide a possible solution directions. In the full level of similarity computing similarity method of calculating the reference to existing answers and answers between the candidates is different: The proposed method is based on the similarity to the reference standard answer, candidates answer calculated relative to a reference answers. This paper elaborates the difference between these two methods, and verify the reasonableness of the proposed method. Finally, design and implement the corresponding comparative experiments and achieved good results, and summed up at one of the shortcomings of the subsequent research work in the future.