Automated Essay Scoring with Deep Learning
Automated assessment of student's assignments is used, e.g., for multiple-choice exams. It is a lot harder for a computer to evaluate essays.
The project will use material from the KAL-project ”Kvalitetssikring av læringsutbyttet i norsk skriftlig”. The aim of this large project was to evaluate the learning outcomes of written Norwegian for students graduating from the basic school (10th grade). The project collected student's exam essays from a four-year period (1998-2001) and analyzed them both quantitatively and qualitatively. The goal was to investigate what the students do – and do not – master, and to which degree the exam grading reflects this.
An earlier master's project has already worked on this task comparing traditional machine learning methods to simple methods based on deep learning. The conclusion from that project was that traditional methods, like linear and logistic regression work best.
The goal of this project is to experiment with more deep learning methods on the same material. e.g. methods for non-linear regression and LSTM, as well as methods combining embeddings and traditional features. Moreover, the project will consider the effect of various splits of the corpus, and possibly the effect of genre.