English exam question answering using a deep learning model

Abstract

In the natural language processing research field, many efforts have been devoted into reading comprehension tasks and deep learning has garnered interests over the recent years, with many different models developed to demonstrate the machine’s ability to compete with humans on the same given task. In order to develop such an intelligent model, a dataset with a high level of reasoning complexity needs to be addressed, along with a well-suited training model which performs significantly well on the given task.

Many studies have been conducted on different reading comprehension tasks and the presence of well-performing models such as Bi-DAF can be seen. However, one of the most challenging tasks, namely the RACE dataset, still has a limited number of studies related to it, due to its nature as a large-scale dataset and intense requirement of complex reasoning. Hence, a detailed study on this dataset can be conducted in order to look into different factors with the purpose of improving the results.

This project aims to improve the result on this particular English reading comprehension task, by choosing the large-scale RACE dataset and improving the baseline model. The study discusses various concepts in layer modeling and attention mechanisms chosen from other well-performing models.

Ultimately, even though the study achieves a relatively acceptable result comparing to some baseline models, it is still relatively behind some newly developed concepts and has many rooms for improvements. A further discussion on the limitations of the developed models can contribute to better results in the future.