Optimizing Features in Active Machine Learning for Complex Qualitative Content Analysis

Abstract:

We propose a semi-automatic approach for content analysis that leverages machine learning (ML) being initially trained on a small set of hand-coded data to perform a first pass in coding, and then have human annotators correct machine annotations in order to produce more examples to retrain the existing model incrementally for better performance. In this “active learning” approach, it is equally important to optimize the creation of the initial ML model given less training data so that the model is able to capture most if not all positive examples, and filter out as many negative examples as possible for human annotators to correct. This paper reports our attempt to optimize the initial ML model through feature exploration in a complex content analysis project that uses a multidimensional coding scheme, and contains codes with sparse positive examples. While different codes respond optimally to different combinations of features, we show that it is possible to create an optimal initial ML model using only a single combination of features for codes with at least 100 positive examples in the gold standard corpus.