Reading comprehension is often tested by measuring a person or system’s ability to answer questions about a given text. Machine reading comprehension datasets have proliferated in recent years, particularly for the English language. The aim of this thesis is to investigate and improve data-driven approaches to automatic reading comprehension. Firstly, I provide a full classification of question and answer types for the reading comprehension task. I also present a systematic overview of English reading comprehension datasets (over 50 datasets). I observe that the majority of questions were created using crowdsourcing and the most popular data source is Wikipedia. There is also a lack of why, when, and where questions. Additionally, I address the question “What makes a dataset difficult?” and highlight the difference between datasets created for people and datasets created for machine reading comprehension. Secondly, focusing on multiple-choice question answering, I propose a computationally light method for answer selection based on string similarities and logistic regression. At the time (December 2017), the proposed approach showed the best performance on two datasets (MovieQA and MCQA: IJCNLP 2017 Shared Task 5 Multi-choice Question Answering in Examinations) outperforming some CNN-based methods. Thirdly, I investigate methods for Boolean Reading Comprehension tasks including the use of Knowledge Graph (KG) information for answering questions. I provide an error analysis of a transformer model’s performance on the BoolQ dataset. This reveals several important issues such as unstable model behaviour and some issues with the dataset itself. Experiments with incorporating knowledge graph information into a baseline transformer model do not show a clear improvement due to a combination of the model’s ability to capture new information, inaccuracies in the knowledge graph, and imprecision in entity linking. Finally, I develop a Boolean Reading Comprehension dataset based on spontaneously user-generated questions and reviews which is extremely close to a real-life question-answering scenario. I provide a classification of question difficulty and establish a transformer-based baseline for the new proposed dataset.