Improving Automatic Sentence-Level Annotation of Human Values Using Augmented Feature Vectors

This paper describes an effort to improve identification of human values that are directly or indirectly invoked within the prepared statements of witnesses before legislative and regulatory hearings. We automatically code human values at the sentence level using supervised machine learning techniques trained on a few thousand annotated sentences. To simulate an actual situation, we treat a quarter of the data as labeled for training and the remaining three quarters of the data as unlabeled for test. We find that augmenting the feature space using a combination of lexical and statistical co-occurrence evidence can yield about a 6% relative improvement in F1 using a Support Vector Machine classifier.