Paraphrase Detection for Short Answer Scoring

We describe a system that grades learner answers in reading comprehension tests in the context of foreign language learning. This task, also known as short answer scoring, essentially requires determining whether a semantic entailment relationship holds between an individual learner answer and a target answer; thus semantic information is a necessary part of any automatic short answer scoring system. At the same time the method must be robust to the particularities of learner language. We propose using paraphrase detection, a method that meets both requirements. The basis for our specific paraphrasing method is word alignment learned from parallel corpora which we create from the available data in the CREG corpus (Corpus for Reading Comprehension Exercises for German). We show the usefulness of this kind of information for the task of short answer scoring. Combining our results with existing approaches we obtain an improvement tendency.