AUTOMATED ESSAY GRADING CONSIDERING DISSIMILARITY FROM INCORRECT RESPONSES

We present a method for grading essays based on Latent Semantic Analysis (LSA), a corpus-based computational method. LSA provides a means of comparing the conceptual similarity between pieces of text without any human intervention. Several systems have successfully used LSA to assess students ’ free-text responses to a professor’s assignment. They usually require LSA to compare students ’ responses with a ‘golden essay’, which corresponds to a standard correct response elaborated by the teacher. As the semantic similarity between the student and golden essay increases, so does the grade for the student’s essay. A potential problem of these systems is that they do not directly deal with incorrect ideas in the student’s essay. Roughly, an incorrect idea will be managed as a low related idea. It will decrease the grade as far as it does not resemble any ideas in the golden essay, but just to the same extent as might do the inclusion of an irrelevant (although not incorrect) sentence. However, in many cases the inclusion of an incorrect idea in a student essay is considered by professors as an index of low comprehension of the topic, which is not necessarily the case with irrelevant ideas. Following this rationale we used LSA to grade student essays, based not only on how much they resemble a golden essay, but also on how much they differ from an essa