System Demo: Tool and Infrastructure for Offensive Language Error Analysis (OLEA) in English

Analysis tool and infrastructure, providing: Such systems excel at detecting and filtering out messages with explicit keywords and mentions; 1. easy-to-use methods for error analysis and however, these systems are known (1) to perform evaluation of new models on existing diagnostic poorly on messages that are implicitly offensive datasets, including model comparison; and that have negation (Röttger et al., 2020; Palmer 2. interfaces to two diagnostic datasets focused et al., 2020); (2) to have annotator biases affecting on nuanced linguistic analysis; and the detection of offensive language (Sap et al., 3. scaffolding to support easy distribution of new 2021); (3) not to be robust to diachronic language datasets and associated analysis methods.