On Releasing Annotator-Level Labels and Information in Datasets

TL;DR

This paper highlights the importance of releasing annotator-level labels in NLP datasets to better capture subjective perceptions and reduce biases caused by aggregating annotations, thereby improving transparency and utility.

Contribution

It demonstrates that label aggregation can obscure individual and group perspectives, and proposes recommendations for more transparent dataset annotation practices.

Findings

Aggregated labels can introduce biases related to socio-cultural backgrounds.

Releasing annotator-level data preserves diversity of perceptions.

Recommendations enhance dataset transparency and utility.

Abstract

A common practice in building NLP datasets, especially using crowd-sourced annotations, involves obtaining multiple annotator judgements on the same data instances, which are then flattened to produce a single "ground truth" label or score, through majority voting, averaging, or adjudication. While these approaches may be appropriate in certain annotation tasks, such aggregations overlook the socially constructed nature of human perceptions that annotations for relatively more subjective tasks are meant to capture. In particular, systematic disagreements between annotators owing to their socio-cultural backgrounds and/or lived experiences are often obfuscated through such aggregations. In this paper, we empirically demonstrate that label aggregation may introduce representational biases of individual and group perspectives. Based on this finding, we propose a set of recommendations for increased utility and transparency of datasets for downstream use cases.