Toward Three-Stage Automation of Annotation for Human Values

Abstract

Prior work on automated annotation of human values has sought to train text classification techniques to label text spans with labels that reflect specific human values such as freedom, justice, or safety. This confounds three tasks: (1) selecting the documents to be labeled, (2) selecting the text spans that express or reflect human values, and (3) assigning labels to those spans. This paper proposes a three-stage model in which separate systems can be optimally trained for each of the three stages. Experiments from the first stage, document selection, indicate that annotation diversity trumps annotation quality, suggesting that when multiple annotators are available, the traditional practice of adjudicating conflicting annotations of the same documents is not as cost effective as an alternative in which each annotator labels different documents. Preliminary results for the second stage, selecting value sentences, indicate that high recall (94%) can be achieved on that task with levels of precision (above 80%) that seem suitable for use as part of a multi-stage annotation pipeline. The annotations created for these experiments are being made freely available, and the content that was annotated is available from commercial sources at modest cost.