Research On Scoring Method Of Short Answer Questions Oriented To Fuzzy Semantics And Multiple Similarities

Short-answer questions,which are subjective questions to evaluate students' ability and accomplishment,are an important type of question on exams.In primary and secondary school,each teacher is responsible for grading lots of exams for different classes by hand.Thus,it is a big challenge for teachers to grade efficiently and fairly.Hence,there are many advantages to use automatic scoring for short-answer questions such as objective fairness,high efficiency,instant feedback,and intelligence analysis.Most of the existing automatic scoring methods for short-answer questions only focus on a single similarity feature and have great limitations in extracting semantic features such as polysemy and context association.Therefore,it is significant to study fuzzy semantic features and multi-similarity methods.The scoring of short-answer questions represents students' skills of reading comprehension,logical thinking,and expressive language.Therefore,it is difficult to use algorithms to score short-answer questions automatically.With the rapid development of artificial intelligence,deep neural networks have shown advantages in text context extraction and word meaning extraction,which provide a new idea for automatic scoring.According to the scoring rules of short-answer questions,this paper deeply studies the relationship between text similarity features,semantic features,shallow features,and scores.These studies will improve the accuracy and universality of automatic scoring for short-answer questions effectively.In order to achieve the goals,this paper uses a new neural network to extract fuzzy semantic features of a text,combines with traditional methods to extract shallow features of the text and introduces a text retrieval probability algorithm.With these methods,this paper calculates the semantic feature similarity,shallow feature similarity,and retrieval probability similarity.Also,it builds an automatic scoring model for short-answer questions using a fusion method.First,this paper studies various existing scoring systems and models according to the scoring rules.And then,it abstracts and extracts attributes of standard answers to short-answer questions and student answers.After that,it extracts scoring indicators and key features which provide a basic basis for subsequent model construction.Secondly,aiming at the limitation of feature extraction in the existing scoring methods and referring to scoring indexes,this paper uses deep learning method to extract semantic features such as word meaning and context relations.Then these features are used to generate sentence vectors,which will be combined into text tensors to calculate the semantic similarity.Thirdly,this paper uses the improved traditional method to extract shallow text features and calculate the similarity of shallow features.Also,a text retrieval probability algorithm is introduced to calculate the similarity of retrieval probability.The scoring model is constructed by combining both the deep learning method and the traditional method.And it combines semantic features and similarity to score.To improve accuracy,two short-answer data sets in Chinese and English from primary and secondary schools were used for experimental analysis.The experimental results show that the scoring method proposed in this paper,which combines fuzzy semantic features with shallow text similarity,deep semantic similarity,and probability similarity,can obtain better semantic feature representation by training sentence vectors with a new bi-directional neural network and a masking language model.The combination of semantic features and similarity features can achieve higher scoring accuracy and a better model universality.