Automatic Dictionary Extraction and Content Analysis Associated with Human Values

Abstract

This paper studies a method for identifying word unigrams and word bigrams that are as- sociated with one or more human values such as freedom or innovation. The key idea is to deterministically associate values with word choices, thus permitting values reflected by sentences to be assigned using dictionary lookup. This approach works nearly as well on average as the most accurate existing methods, but the principal contribution of the new method is that the basis for the system’s classification decisions are more easily interpreted by social scientists. The new method is based on using a Monte Carlo algorithm with sim- ulated annealing to efficiently explore the space for optimal assignments of human values to unigrams and bigrams. Results are reported on an annotated test collection of prepared statements from witnesses at public hearings on the topic of net neutrality. The results include both accuracy comparisons with a previously reported approach and the use of emergent human coding to explain the classification process in a way that social scientists find to be useful as a way of characterizing the use of word pairs to express human values in this context.