Sentiment is Not Stance: Target-Aware Opinion Classification for Political Text Analysis

Abstract Sentiment analysis techniques have a long history in natural language processing and become standard tool the of political texts, promising conceptually straightforward automated method extracting meaning from textual data by scoring documents on scale positive to negative. However, while these kinds sentiment scores can capture overall tone document, underlying concept interest for is often actually document’s stance with respect given target—how positively or negatively it frames specific idea, individual, group—as this reflects author’s attitudes. In paper, we question validity approximating author through advocate greater attention be paid conceptual distinction between its stance. Using examples open-ended survey responses discussions social media, demonstrate that many text applications, do not necessarily align, therefore methods fail reliably ground-truth document stance, amplifying noise leading faulty conclusions.