Modeling Human Perspectives with Socio-Demographic Representations

ACL ARR 2026 January Submission718 Authors

Keywords: human annotation variation, annotation disagreement, human perspectives

Abstract:

Recent studies show that many NLP tasks contain instances with annotation disagreement and multiple valid perspectives.
Modeling these perspectives and understanding their association with socio-demographic attributes has also received growing attention.
In this work, we propose a novel architecture that jointly models annotator perspectives and learns annotators' socio-demographic representations from their annotation patterns. Our approach not only improves the performance of hate speech and toxic content prediction but also produces meaningful annotator representations. These representations enable further analysis and visualization of the relationships between socio-demographic attributes and variations in annotators’ perspectives.

Paper Type: Long

Research Area: Computational Social Science, Cultural Analytics, and NLP for Social Good

Research Area Keywords: hate-speech detection,human behavior analysis

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 718