Tagging Reading Comprehension Materials With Document Extraction Attention Networks

Abstract

Reading comprehension tasks are commonly used for developing students' reading ability. In order to adaptively recommend reading comprehension materials to students engaged in computerized testing, the information in an item bank (a collection of test items stored in a dataset) must be effectively indexed. Familiarity with the topics present in the documents influences students' reading performance. As different question types require different skills, we tag documents with topics and questions with their corresponding types to measure the students' abilities and subsequently recommend relevant materials to them. However, automatic tagging has not been extensively studied in this field. In this article, we propose a document extraction attention network (DEAN) to accomplish the two aforementioned tasks. For topic tagging, DEAN utilizes questions to increase the sample size of documents implicitly through multitask learning. For type tagging, DEAN leverages the information gathered from documents, which aids in the task of prediction. Experiments demonstrate the effectiveness of our mutual use of information obtained from documents and questions. Results indicate that DEAN outperforms commonly used text classification methods when tested on a reading comprehension dataset.