Semi-automatic analysis of traditional media with machine learning

Abstract

The analysis of traditional and social media is a non-trivial task, requiring the input of human analysts for quality. However, the ready availability of electronic resources has led to a large increase in the amounts of such data to be analysed: the quantities of data (tens of thousands of documents per day) mean that the task becomes too substantial for human analysts to perform in reasonable time frames and with good quality control. In this project, we have explored the use of machine-learning techniques to automate elements of this analysis process in a large media-analysis company. Our classifiers perform in the range of 60%-90%, where an average agreement between human analysts is around 80%. In this paper, we examine the effect of using active-learning techniques to attempt to reduce the amount of data requiring manual analysis, whilst preserving overall accuracy of the system. © Springer-Verlag London Limited 2011.