Not All Subjectivity Is the Same! Defining Desiderata for the Evaluation of Subjectivity in NLP

1. Problem Statement

Natural Language Processing (NLP) models are increasingly deployed in societal contexts involving subjective concepts (e.g., hate speech detection, moral reasoning, value alignment). However, current evaluation practices often fail to account for the nuances of subjectivity.

• The "Majority Collapse": Standard training paradigms (e.g., RLHF, DPO) and evaluation metrics often aggregate diverse human judgments into a single "gold" label (majority vote). This collapses minority perspectives and obscures valid disagreements, leading to models that reflect dominant viewpoints while marginalizing others.
• Lack of Granularity: Existing literature treats subjectivity as a monolithic concept, failing to distinguish between different types of subjective inputs (e.g., ambiguity vs. genuine polyphony) or the sources of disagreement (task definition, input context, or annotator demographics).
• Evaluation Gap: There is no standardized framework to evaluate whether a model correctly identifies, reasons about, and expresses subjective diversity without hindering user experience.

2. Methodology

The authors propose a top-down approach to define evaluation criteria, followed by a bottom-up empirical analysis of current practices.

A. Conceptual Framework

The paper first deconstructs subjectivity in data and models:

1. Sources of Subjectivity: Drawing on the "Triangle of Reference," they identify three sources:
   • Task Ambiguity: Imprecise guidelines.
   • Input Ambiguity: Lack of context (e.g., reclaimed slurs).
   • Annotator Attitude: Demographics, experiences, and emotional states.
2. Dimensions of Input: The authors introduce a 2x2 matrix to characterize inputs:
   • Ambiguity vs. Non-Ambiguity: Whether the input requires disambiguation (e.g., "take this bitch out" – dog vs. person).
   • Polyphony vs. Monophony: Whether multiple valid perspectives exist (polyphony) or a single ground truth exists (monophony).
   • Key Insight: Models must handle these dimensions differently. For example, an ambiguous input requires clarification, while a polyphonic input requires representing multiple valid viewpoints.

B. Proposed Desiderata (The Core Contribution)

Based on the framework, the authors propose seven sequential desiderata for evaluating subjectivity-sensitive models:

1. D1: Recognize when: Identify if an input requires subjective modeling (vs. objective fact).
2. D2: Recognize which: Identify the type of subjectivity (ambiguity, polyphony, or both).
3. D3: Recognize why: Explain the source of the subjectivity (e.g., missing context vs. demographic difference).
4. D4: Predict: Accurately predict the distribution of human judgments or specific perspectives.
5. D5: Calibrate: Ensure the model's confidence/distribution aligns with the actual frequency of human opinions.
6. D6: Represent: Explicitly capture minority voices and diverse perspectives, avoiding bias toward dominant views.
7. D7: Express: Communicate these perspectives to the user effectively without degrading the user experience (e.g., avoiding overwhelming lists of every possible view).

C. Empirical Analysis

To validate the relevance of these desiderata, the authors conducted a systematic review of 60 NLP papers dealing with subjectivity. They mapped existing evaluation metrics (e.g., Accuracy, F1, KL-Divergence, Inter-Annotator Agreement) to the seven desiderata.

3. Key Contributions

1. Distinction between Ambiguity and Polyphony: The paper rigorously defines and separates "ambiguity" (lack of information requiring disambiguation) from "polyphony" (co-existing valid perspectives). This distinction is argued to be critical for designing appropriate model behaviors.
2. Seven-Step Evaluation Framework: The introduction of the sequential desiderata (D1–D7) provides a comprehensive checklist for developing and evaluating models that handle subjectivity, moving beyond simple accuracy metrics.
3. Systematic Gap Analysis: The paper provides a quantitative mapping of the current literature, revealing that:
   • Most research focuses on D4 (Predict) and D6 (Represent).
   • D2 (Recognize which) and D3 (Recognize why) are significantly understudied.
   • D7 (Express) is almost entirely ignored; no papers evaluated how subjectivity is communicated to users.
   • Current metrics rarely address the interplay between desiderata (e.g., how failing at D3 impacts D4).

4. Results and Findings

• Metric Limitations: Current evaluation relies heavily on Distance Metrics (KL-divergence, Wasserstein) for distributional alignment and Accuracy/F1 for single-label aggregation. These are insufficient for evaluating the reasoning behind a model's subjective output.
• Understudied Areas:
  • D2 & D3: Very few papers explicitly evaluate whether a model can distinguish between an ambiguous input and a polyphonic one, or explain why a disagreement exists.
  • D7 (Expression): There is a complete lack of metrics for evaluating the usability of subjective outputs. Models may generate correct distributions but present them in ways that confuse or overwhelm users.
• Data Gaps: Existing datasets often lack the necessary granularity (e.g., annotator demographics, confidence scores, or raw disagreement data) to support the full evaluation of these desiderata.

5. Significance and Future Directions

• Real-World Impact: The paper argues that for NLP models to be safe and equitable in real-world deployment, they must not just "predict" the majority view but actively represent minority perspectives and explain their reasoning.
• Interdisciplinary Collaboration: The authors call for integrating insights from Human-Computer Interaction (HCI), Affective Computing, and Social Science to develop better evaluation strategies, particularly for D7 (Expression) and dataset construction.
• New Research Avenues:
  • Developing metrics for uncertainty estimation to distinguish between aleatoric (input) and epistemic (model) uncertainty.
  • Creating benchmarks that test the entanglement of ambiguity and polyphony.
  • Designing explainability techniques that verify if a model is making predictions for the "right reasons" (D3).

In summary, this paper serves as a foundational position paper arguing that "subjectivity" is not a single problem to be solved by better accuracy, but a multi-faceted challenge requiring a new suite of evaluation criteria that prioritize the representation, reasoning, and communication of diverse human perspectives.