EVALUATING SEMANTIC ANALYSIS METHODS FOR SHORT ANSWER GRADING USING LINEAR REGRESSION

The assessment of free-text answers may demand significant human effort, especially in scenarios with many students. This paper focuses on the automatic grading of short answer written in Portuguese language using techniques of natural language processing and semantic analysis. A previous study found that a similarity scoring model might be more suitable to a question type than to another. In this study, we combine latent semantic analysis (LSA) and a WordNet path-based similarity method using linear regression to predict scores for 76 short answers to three questions written by high school students. The predicted scores compared well to human scores and the use of combined similarity scores showed an improvement in overall results in relation to a previous study on the same corpus. The presented approach may be used to support the automatic grading of short answer using supervised machine learning to weight different similarity scoring models.