Capturing Perspectives of Crowdsourced Annotators in Subjective Learning Tasks

TL;DR

This paper introduces AART, a novel method that learns annotator representations to better capture individual perspectives in subjective classification tasks, addressing biases from traditional label aggregation.

Contribution

The paper proposes AART, a new approach that models annotator behaviors to improve subjective classification and fairness, especially in crowdsourced datasets.

Findings

AART improves performance in capturing individual annotator perspectives.

The method enhances fairness for marginalized annotators.

Experimental results show better alignment with annotator-specific labels.

Abstract

Supervised classification heavily depends on datasets annotated by humans. However, in subjective tasks such as toxicity classification, these annotations often exhibit low agreement among raters. Annotations have commonly been aggregated by employing methods like majority voting to determine a single ground truth label. In subjective tasks, aggregating labels will result in biased labeling and, consequently, biased models that can overlook minority opinions. Previous studies have shed light on the pitfalls of label aggregation and have introduced a handful of practical approaches to tackle this issue. Recently proposed multi-annotator models, which predict labels individually per annotator, are vulnerable to under-determination for annotators with few samples. This problem is exacerbated in crowdsourced datasets. In this work, we propose \textbf{Annotator Aware Representations for Texts (AART)} for subjective classification tasks. Our approach involves learning representations of annotators, allowing for exploration of annotation behaviors. We show the improvement of our method on metrics that assess the performance on capturing individual annotators' perspectives. Additionally, we demonstrate fairness metrics to evaluate our model's equability of performance for marginalized annotators compared to others.