Data driven methods for analysis and improvement of academic English writing exercises

Abstract

Technology assisted learning and teaching has become an integral part of university education. Online learning systems are especially useful for language learning as they allow students to better practice the material and study at their own pace. These systems enable teachers to reach more students and reduce workload needed to correct student work. They become particularly helpful in the case of writing exercises. This thesis presents the process of improving an online learning content distribution system, the Acos server, from both student and teacher perspectives with the help of data driven methods. The work is focused on a concrete exercise type provided by the system, short answer question type, in the context of learning academic English. The thesis is divided into two parts: implementation of new user interface features of the system and log data analysis of past student activity using the Acos server.

The newly implemented system features include a spellchecker, detailed feedback functionality and hint button functionality. The features led to improvement of student performance in solving the questions and decreased the difficulty level of the exercises.

The proposed clustering approach for students' answers analysis aims to discover certain patterns in student behaviour and better identify their mistakes. The clustering is visualized using an interactive interface. Different feature extraction methods are compared based on both sentence syntactic and semantic structure. Constituency and Wordnet based features yield the best results for syntax based and semantics based clustering of student answers, respectively.