Voices in a Crowd: Searching for Clusters of Unique Perspectives

TL;DR

The paper addresses the challenge of reproducing majority biases in language models by uncovering minority perspectives through annotator-behavioral embeddings rather than relying on annotator metadata. It introduces a two-stage framework: a supervised component that predicts individual annotator labels to form behavioural embeddings, and an unsupervised component that reduces dimensionality and clusters these embeddings to identify voices, validated by internal and external metrics and qualitative case studies on MBIC and GWSD datasets. The approach reveals three voice types—majority, minority, and inter-minority—demonstrating robustness and generalization across datasets while capturing nuanced intersectional perspectives. This work advances bias-aware modeling by enabling dynamic, metadata-free discovery of diverse viewpoints, with implications for more inclusive and auditing-friendly AI systems.

Abstract

Language models have been shown to reproduce underlying biases existing in their training data, which is the majority perspective by default. Proposed solutions aim to capture minority perspectives by either modelling annotator disagreements or grouping annotators based on shared metadata, both of which face significant challenges. We propose a framework that trains models without encoding annotator metadata, extracts latent embeddings informed by annotator behaviour, and creates clusters of similar opinions, that we refer to as voices. Resulting clusters are validated post-hoc via internal and external quantitative metrics, as well a qualitative analysis to identify the type of voice that each cluster represents. Our results demonstrate the strong generalisation capability of our framework, indicated by resulting clusters being adequately robust, while also capturing minority perspectives based on different demographic factors throughout two distinct datasets.

Figures (2)

• Figure 1: Different approaches for handling annotations: i) disagreement-based create per-example distributional labels which fail to account for dataset-level effects; ii) metadata-based train models on annotations linked with annotator metadata, which often groups disagreeing annotators who share metadata labels; iii) the "Voices in a crowd" approach dynamically creates clusters based on annotation patterns and finally verifies each cluster as a voice based on post-hoc matched metadata labels.
• Figure 2: Training component: 6 modelling architectures for extracting hidden states (denoted with a yellow circle as $Emb_n$) used as input for the Clustering component.