This is the repository for my Master thesis research: "Measuring readability of technical texts".

Abstract: This research explores various approaches to measuring readability of technical texts. The data I work with is provided by Hyperskill, an online educational platform dedicated mostly to Computer Science, where I did my internship. In the first part of my research, I examine classical readability formulas and try to find correlations between their values and the user statistics available for the texts. The results show that there are no high correlations, thus, the standard formulas are not suitable for the task. The second part of the research is dedicated to experiments with machine learning algorithms. Firstly, I use four sets of features to predict the average rating, completion time, and completion rate of a step. Then, I introduce a rule-based algorithm to split the texts into well- and poorly-written ones, which relies on students’ comments. However, binary classification trained on this division shows low results and is not used in the final pipeline. The system suggested as the outcome of my work employs the user statistics’ prediction for new texts and a rule-based comment-focused algorithm for the published texts. As a result, texts that the system identifies as “bad” are reported to the Hyperskill content team for improvement.

The repository contains Jupiter Notebooks with code implementing various steps of my work. In general, they can be groupsed into four broad parts:

• Data exploration and preparation: notebooks 1-3
• Experimets with correlations: notebooks 4-7
• Work with comments: notebooks 8 and all from 9
• Machine learning experiments: notebooks 10-13

Each notebook contains headings describing the main steps of the work, as well as some comments. The datasets used in the research are not given in the repo since they were provided by Hyperskill and are protected by NDA.