An affordance-inspired tool for automated web page labeling and classification

Abstract

Writing programs that are capable of completing complex tasks on web pages is difficult due to the inconsistent nature of the pages themselves. While there exist best practices for developing naming schemes for page elements, these schemes are not strictly enforced, making it difficult to develop a general-use automated system. Many pages must be hand-labeled if they are to be incorporated into an automated testing framework. In this thesis, I build an application that assists human users in classifying and labeling web pages. This system uses a gradient boosting classifier from the scikit-learn Python package to identify which of four tasks may be performed on a given web page. It also attempts to automatically label the input fields and buttons on the web page using a gradient boosting classifier. It outputs its results in a format that can be easily consumed by the LARIAT system at MIT Lincoln Laboratory, greatly reducing the human labor required to incorporate new web pages into the system.