Detection of Incorrect Case Assignments in Paraphrase Generation

This paper addresses the issue of post-transfer process in paraphrasing. Our previous investigation into transfer errors revealed that case assignment tends to be incorrect, irrespective of the types of transfer in lexical and structural paraphrasing of Japanese sentences [3]. Motivated by this observation, we propose an empirical method to detect incorrect case assignments. Our error detection model combines two error detection models that are separately trained on a large collection of positive examples and a small collection of manually labeled negative examples. Experimental results show that our combined model significantly enhances the baseline model which is trained only on positive examples. We also propose a selective sampling scheme to reduce the cost of collecting negative examples, and confirm the effectiveness in the error detection task.