In traditional machine learning, annotator disagreement is treated as a failure. If three people label a sentence as "toxic" and two label it as "neutral," the standard procedure is to take the majority vote and discard the minority view as noise. This "monistic" approach assumes there is a single, objective ground truth—a "gold standard"—that the model must uncover.
However, for subjective tasks like irony detection, humor, or hate speech, there is often no single truth. As Cecilia Ovesdotter Alm argued in Subjective Natural Language Problems, treating disagreement as noise ignores the nuanced, social reality of human communication. A new research direction, often called Data Perspectivism, suggests that annotator variance is not an error to be filtered, but a vital signal to be modeled.
The Signal in the Noise
When annotators disagree on whether a comment is offensive, the disagreement itself tells us something about the text. A high-variance label suggests the content is ambiguous, controversial, or culturally sensitive. Alastair J. D. Craft and colleagues demonstrated this in the context of music genre classification, noting that labels are often "cultural responses" rather than inherent properties of the data (How Many Beans Make Five?).
By preserving these disagreements, models can learn the distribution of human opinions rather than just the average. This is the core of the LeWiDi (Learning With Disagreements) framework, which challenges AI to predict the full spectrum of human judgments. As Aneesha Sampath and Victoria Lin showed with SeedBERT, recovering these distributions leads to more robust performance in affective computing because the model no longer tries to force a single pattern onto data that naturally supports multiple interpretations.
Architectural Strategies for Subjectivity
How do we build models that respect multiple viewpoints? Researchers have converged on several key mechanisms:
- Soft-Label Multi-Task Learning: Instead of training a model to predict a binary 0 or 1, researchers like Tommaso Fornaciari and Barbara Plank use multi-task frameworks to predict both the majority label and the distribution of annotator opinions (Beyond Black & White). This forces the model to represent the inherent uncertainty of the task.
- Perspectivist Adaptation: Rather than one global model, we can create models that adapt to specific viewpoints. Daniil Ignatev and Massimo Poesio introduced a hypernetwork approach that generates custom weights (LoRA adapters) for individual annotators. This allows a single system to simulate thousands of different perspectives efficiently.
- Demographic Awareness: Disagreement is often structured by who the annotators are. Eve Fleisig and Dan Klein found that modeling annotator demographics allows systems to prioritize the perspectives of groups most affected by specific content, such as targeting offensive speech detection toward the communities it harms (When the Majority is Wrong).
From Consensus to Pluralism
The shift toward perspectivism has profound implications for AI alignment. Traditional Reinforcement Learning from Human Feedback (RLHF) often seeks a "monistic" alignment—tuning a model to a single average preference. Taylor Sorensen and Jared Moore argue instead for Pluralistic Alignment, where models are designed to represent the diverse and often conflicting values of the human population.
This transition is visible in specialized datasets like EPIC, which preserves individual irony ratings across different ages and nationalities, and RAVE, which captures how different legal experts prioritize different facts. As Julia Romberg notes in her call for a perspectivist turn in argument quality, the goal is no longer to build a machine that decides what is "good" or "true," but one that understands for whom an argument is persuasive or a joke is funny.
By treating disagreement as a structured signal, perspectivist modeling moves AI away from sterile averages and toward a more nuanced, human-centric understanding of language.