Annotator Disagreement

Depression has no diagnostic blood test. Language models promise tireless, consistent assessment, but can accurate raters disagree about individuals? We pre-registered 880 language-model raters, crossing 11 open models with prompting and scoring choices, and applied them to 189 interviews against the eight-item Patient Health Questionnaire. Model choice explained 30.0% of summed-symptom score variance, stable participant differences 10.5%. Two randomly drawn raters with area under the receiver operating characteristic curve (AUC) >= 0.70 disagreed on screening decisions for 40% of participants, on average. Average over-rating governed how many were flagged, yet equal-capacity raters chose differently for about one participant in five. A locked analysis of 86 new interviews reproduced the main pre-registered findings. Exploratory recalibration with 40 labelled participants raised accuracy from about 60% to 75% and halved disagreement, leaving one participant in five decided differently. Calibration repaired much of the rater dependence without securing agreement about individuals.

How much of a conference accept/reject decision would change if the same paper were reviewed by a different set of reviewers? Running a second independent program committee is the gold standard for answering this, but it is prohibitively expensive: done only twice (NeurIPS 2014 and 2021). We build an observational estimator of this quantity from public review data alone, calibrate it twice, and apply it to nine years of ICLR (2017-2025; 36,113 papers, 134,912 reviews). The estimator decomposes scores with a Bayesian ordered-probit model into paper quality and reviewer noise, maps scores to decisions with a logistic model, and simulates two independent committees (posterior draws B=1,000; committee sizes k=2,3,4). Estimated disagreement rates are 23-30% at k=2 and 18-24% at k=4; 30-50% of accepted papers would be rejected. External calibration: at the NeurIPS 2021 reviewer-count caliber (k=3), the simulated 2021 disagreement rate is 23.3% [21.7%, 25.0%] vs. reported 23.0% (bias +0.3pp); accept precision and committee correlation agree within 5pp and 0.04. Internal calibration: on 18,740 papers with 4+ reviews, random model-free 2+2 reviewer splits agree with the k=2 simulation within 1pp in 2018 and 2021-2025. Longitudinally, we find no robust time trend in reviewer noise over 2017-2025. The high accepted-paper flip rates of 2020 and 2021 have distinct mechanisms: the 2020 four-point scale compressed scores (23.7% of papers had zero within-paper variance), and a counterfactual shows coarsening the scale raises disagreement by about 7pp; 2021 instead combined the lowest signal-to-noise ratio in the sample with the most threshold-crowded acceptances. For the LLM era, a 2023 breakpoint test on within-paper score variance finds no break, but the design has almost no power, and no post-2022 review text or confidence data exist, so no LLM attribution is attempted.

Structured span annotations, such as quantities with their units, uncertainty modifiers, and event classes, are expensive to create and hard to keep trustworthy once language models enter the loop. We present QuanReview, an open-source system for auditing and correcting such annotation layers. QuanReview aligns two annotation streams over the same documents at character level, resolves unambiguous cases by an explicit and logged policy, and routes candidate conflicts to a browser-based adjudication interface where reviewers accept either side, build field-level hybrids, or flag items for re-annotation. A campaign manager assigns documents to multiple annotators with configurable redundancy, computes agreement at document and span level, auto-merges unanimous documents, and exports the corrected layer in the original file format, so that it can replace the original annotation files directly. Applied to a 4,457-record humanitarian benchmark and an LLM extraction stream, the system fully auto-merged 8% of documents, applied automatic policy decisions to a further 1,513 records, and concentrated human attention on 3,131 candidate conflicts, a mean of 5.4 per reviewed document.

In automated clinical coding, where the label space spans tens of thousands of diagnosis and procedure codes, models are currently evaluated against a single gold annotation, treating any deviation as error. But we find when two teams code the same 110 ACI-Bench encounters, they agree on only 73% of codes (Jaccard similarity) for the same note; even after an independent clinical audit removes erroneous codes, agreement rises only to 77%. Is that gap error or something systematic? We model the systematic component as coding style $ψ$, a coder- or site-specific policy over what to code and how much to document, and recast coding as $p(\mathrm{code}\mid\mathrm{note},ψ)$, estimating $ψ$ with a 10-dimension rubric. If style were noise, conditioning on it would do nothing. Instead, across five datasets a model conditioned with a data-matching style raises ICD F1 by up to 26 points and an extreme mismatched one lowers it by up to 21. Four prompt based coding methods spanning 39-49 F1 converge to 52-56 once style is supplied (All p<0.05). Much of what single-gold evaluation charges to model error is recoverable, unmodeled style.

A panel's human-equivalent size is target-specific. Matching a fixed 32-judge panel to empirical human label distributions on three ChaosNLI tasks yields two distinct effective sizes: distributional-error matching gives $ν_{\mathrm{MSE}}=2.304$, $3.750$, and $3.445$, whereas spectral matching gives $ν_H=4.242$, $6.459$, and $6.499$, a gap of $1.72$--$1.89\times$; a binary-error diagnostic credits the same panels with only $1.971$--$2.227$ effective votes. Extrapolating the distributional-error curve at fixed squared mean residual, mean member variance, and normalized mean covariance gives asymptotes of $2.392$, $3.990$, and $3.655$, with 32 judges already reaching $94.0$--$96.3\%$. An exact spectral identity explains the gap: error depends on member energy and on the orientation of residual variation relative to averaging, information that the participation ratio (PR) discards. A realizable hard-label construction confirms that higher spectral diversity can coexist with worse distribution recovery even under equal member energies and nonnegative correlations, and the consensus direction retains $γ_{\mathrm{co}}=43.8\%$, $33.7\%$, and $35.9\%$ of centered residual variance. An external check on CC-1000, a 1,000-item Civil Comments subset with a different panel, gives $ν_H=2.84$. For panel choice, we establish an existence result and one feasible path: exhaustive enumeration at $k\in\{5,7\}$ shows that panels beating the accuracy-top-$k$ baseline on both accuracy and $ν_H$ always exist, and greedily swapping at most two members reaches $24.8$--$56.0\%$ higher $ν_H$ at $0.10$--$1.10$ percentage points higher accuracy. Our dataset and code are available at https://github.com/Chao1208/32judges-votes.

Dynamic facial expression recognition (DFER) benchmarks such as DFEW provide multiple annotator votes per clip, yet most models collapse them to a majority label and cannot represent human disagreement at inference time. We propose a disagreement-aware DFER framework that trains directly on the raw annotator count vector using a Dirichlet-Multinomial likelihood. Unlike mean-only soft-label objectives, the proposed likelihood provides scale-sensitive supervision for the Dirichlet concentration while preserving the predictive mean. A separate ambiguity head predicts annotation entropy for unseen clips, and a monotone Chow-style reject rule combines predicted ambiguity, vacuity, temporal instability, and input quality for selective prediction. On DFEW, the method preserves recognition accuracy while reducing ECE by 30% and AURC by 15%, and predicted ambiguity reaches a Spearman correlation of 0.52 with the annotation entropy of test clips. The calibration and selective-prediction gains transfer to FERV39k and remain under identity- and movie-disjoint DFEW splits.

Sea-ice stage of development (SoD) describes the age and associated thickness of sea ice and provides important information for navigation, and operational ice monitoring. SoD labels are obtained from operational ice charts, where trained analysts interpret satellite observations and assign standardized stage codes to regions with similar ice conditions. These codes often represent ranges of compatible ice thicknesses rather than exact physical values. Deep-learning methods can automate SoD mapping and commonly adopt operational ice charts as reference labels for training. These annotations are not exact, however; this is because chart interpretation relies on analyst judgement and on the observations available at the time, so different ice services may assign different SoD labels to the same conditions. We term this variation across independently produced expert annotations multi-annotator label uncertainty; collapsing the annotations into a single deterministic target discards this variation. A second source of uncertainty originates in the learned model itself. In this paper, we quantify both sources: annotation uncertainty from disagreement among independent ice-service charts and model uncertainty from the learned predictive models. We then evaluate their relationship by testing whether model uncertainty is higher where ice services disagree. We observe that supervision incorporating information from multiple annotators can improve this correspondence, with soft supervision achieving the highest overall correlation of 0.256. The relationship becomes substantially stronger near the ice edge, where model predictive uncertainty closely tracks multi-annotator disagreement, reaching a correlation of 0.704 within 0--10 km. Among the uncertainty-estimation approaches, Monte Carlo dropout provides the best-calibrated confidence estimates, with an expected calibration error of 0.050.

Natural-language tasks can elicit different verdicts from protocol-following evaluators that receive the same declared information. We study aggregate disambiguation systems (ADSs). Given a task and a candidate solution, each evaluator casts a binary vote on whether the solution should be accepted, and the system aggregates the votes of a finite panel. The target is protocol reproducibility relative to an explicitly declared evaluator reference, not semantic truth. We separate fixed finite censuses, probabilistic evaluator populations, and growing-census limits, since their endpoint laws and guarantees are not interchangeable. In the population setting, we use finite samples to estimate how often a finite panel reaches the same decision as the declared evaluator population. We provide a lower confidence bound on the fraction of candidate solutions for which the disagreement probability is at most a chosen tolerance. The calculation accounts separately for sampling candidate solutions and sampling evaluators. The construction permits arbitrary dependence among columns induced by shared evaluator rows and uses exact binomial intervals at the evaluator layer and an exact one-sided binomial inversion at the generator layer. Simulations check the implementation against known population coverages and expose power limitations.

Media bias detection relies on definitions and examples that specify what counts as bias, yet these specifications often vary across datasets or remain implicit, even when given the same name. Such variation makes it unclear whether models trained for the same bias category learn the same construct or different phenomena, a problem largely overlooked in prior work. We examine how definition choice affects bias annotation in a between-subjects experiment with 354 participants and a parallel evaluation with four LLMs. Participants and models rate six news articles across four bias categories using definitions that vary in conceptual framing and elaboration. Across 8,496 human and 28,800 LLM ratings, we find that the conceptual target of a definition drives annotation divergence, while construct-preserving elaboration does not: conceptual framing significantly shifts annotations for humans and does so even more strongly for LLMs. We discuss implications for construct specification in annotation protocols and prompt-based measurement, and consider how definitional sensitivity may propagate to downstream classification beyond media bias. We also release MUDD, the Multi-Definition Bias Detection Dataset.

When people label text for sexism, they often disagree, and not because some of them are wrong: they genuinely perceive sexism differently. Most NLP systems discard this disagreement by collapsing it into a majority vote. We propose the Multi-Agent Perspectivist Preference Optimization (MAP-PO) framework to keep these different perspectives. On the EXIST 2024 dataset of labeled English and Spanish tweets, we first cluster annotators by their labeling behavior rather than their demographic attributes. We then fine-tune one Large Language Model agent per cluster to reproduce that cluster's annotation behavior, and coordinate the agents with preference optimization that combines individual and team-level rewards. We evaluate MAP-PO in four settings defined by two languages and two backbone language models, asking whether each agent reproduces the annotations of its own cluster and whether the agents together reproduce the majority label. Two findings hold in all four settings. First, without fine-tuning the agents behave almost identically, so cluster-specific training is necessary. Second, we show that training each agent only on the labels of its own cluster pushes the agents far beyond the clusters they should represent, while adding a shared team-level training signal consistently keeps each agent calibrated to its cluster.

Prior work on human label variation (HLV) in natural language inference (NLI) has often relied on re-annotation resources that select items by disagreement level. An earlier study (arXiv:2607.15870) found that hypotheses containing non-upward monotonicity operators showed lower label agreement in ChaosNLI (Cliff's delta = -0.284), which is restricted to items whose majority label carries exactly three of five votes. We preregistered a replication of this boundary in the unselected populations that ChaosNLI was drawn from: the SNLI and MultiNLI development sets, using the same operator tagger and a four-level ordinal agreement outcome. The registered prediction fails. All seven contrasts return a positive Cliff's delta (non-upward items agree slightly more, not less), the only significant confirmatory contrast has the opposite sign to the registration, and every effect is far below our smallest effect size of interest (0.10). Robustness checks support the measurement: simulated tagger misclassification shrinks the effects rather than manufacturing them, and a manual re-tagging audit reaches four-class agreement of 0.875 on a fresh 200-item sample. We conclude that the earlier negative boundary is plausibly a structure conditional on low-agreement selection rather than a population-level property, and that HLV structure claims built on selected re-annotation resources should state their selection conditional explicitly.

Human label variation in natural language inference (NLI) is increasingly treated as a signal to be measured rather than as noise to be removed. We ask whether one candidate source of that signal, formal semantic structure such as negation, quantification, and monotonicity, changes how much annotators disagree and what they disagree about. We tag the SNLI and MNLI items of ChaosNLI, each labeled by 100 annotators, with a rule-based monotonicity tagger, check the tagger by hand on a sample of the same items, and answer three questions. At the group level, hypotheses that are not purely upward monotone attract somewhat more disagreement, but an error sensitivity analysis shows that this difference is sensitive to tagger error. At the item level, formal structure explains only a few percent of the variation and cannot pick out the items that attract high disagreement. In composition, the kinds of disagreement recorded by VariErr and LiTEx do not differ detectably across the formal boundary. Formal structure therefore belongs in the inventory of disagreement sources, with a small and bounded weight. ChaosNLI was built from low-agreement items, and every claim holds within that scope. Analysis decisions were written in a version-controlled research log before the corresponding results were computed, and negative results are reported in full.

Subjective NLP tasks often exhibit systematic annotator disagreement, requiring models that represent uncertainty rather than collapse it. We introduce Ensemble Diversity Optimization (EDO), a prediction-space framework that jointly optimizes ensemble weights, effective cardinality, and calibration through a unified differentiable objective. EDO learns ensemble composition and size end-to-end via Gumbel-Softmax relaxation and incorporates a signed diversity regularizer, tuned on validation data, to steer optimization toward either preserving or suppressing disagreement. This regularization prevents ensemble collapse and enables controlled navigation of the utility-calibration trade-off. The framework integrates a soft F1 surrogate, class-weighted cross-entropy to address imbalance, and reliability-weighted diversity to regulate intra-ensemble variability. Experiments on four subjective text-classification benchmarks (ArMIS, ConvAbuse, HS-Brexit, MD-Agreement) show that EDO substantially improves probabilistic calibration, reducing cross-entropy (40-78% depending on baseline) and lowering Brier scores relative to Soft-CE, Soft-MD, Top-5 Voting, and WEL, while maintaining competitive F1 and better alignment with annotator distributions. These results demonstrate that jointly optimizing ensemble structure with a signed diversity regularizer provides an efficient, model-agnostic approach for modeling human subjectivity in supervised learning.

Emotion recognition of song lyrics is a challenging task since lyrics may not necessarily align with the overall emotion of a song. As a result, lyrics annotation remains largely underexplored. Drawing inspiration from research in large language model (LLM) assisted annotation, we examine the alignment between humans and LLMs for annotation of lyrics by creating a new sentence-level dataset of lyrics. Our observations highlight the subjectivity of the task and the inherent challenges. Following this, we present a hybrid annotation framework that optimizes human and LLM annotation by predicting potential misalignment in annotation.

Hate speech annotation pipelines routinely collapse annotator disagreement into majority vote labels before training. We show that this aggregation is not neutral: 42.6% of all annotator disagreement in HateXplain concentrates specifically at the hate/offensive boundary, a pattern consistent with annotators applying different thresholds for where hate begins (chi-squared = 135.199, df = 2, p < 0.0001). Both a hard-label BERT model (Model A) and a soft-label model (Model B) drop 22 percentage points in accuracy from agreed posts (~80%) to disagreement posts (~58%), confirmed at p < 0.0001. A per-annotator multi-head model (Model C) widens this gap further to 28 points while collapsing offensive disagreement accuracy to 0.245. Critically, Model A expresses significantly higher confidence on boundary case errors than Model C (0.710 vs. 0.495, p < 0.0001), meaning standard evaluation metrics will not detect the failure. Three downstream interventions of increasing sophistication all fail to recover boundary accuracy. We argue the problem is structural. Majority vote presents a contested judgment as ground truth, and models inherit that false certainty. The intervention must be upstream in annotation design.

Speech emotion recognition (SER) often relies on hard consensus labels that collapse annotator disagreement. We study distribution-based supervision for 9-class SER on MSP-Podcast 2.0 using a WavLM-Base multitask model for categorical emotion and dimensional VAD. Hard-label training is compared with targets from primary and merged primary--secondary annotator vote distributions. Distributional objectives improve alignment with human vote distributions, reducing JSD/KLD relative to hard-label training. Analysis shows that hard supervision partly benefits from assigning ambiguous utterances to the residual Other class, whereas distributional supervision redistributes uncertainty across emotion categories. Entropy-stratified evaluation shows that high-ambiguity utterances remain challenging, but distribution-based supervision better captures perceptual uncertainty. These findings support moving beyond hard labels toward targets that reflect listener disagreement.

As previous research on annotator disagreement in discourse phenomena has shown, understanding text coherence varies considerably from one individual to another. To explore this phenomenon, we created two corpora with multiple annotations of Czech texts, accompanied by annotators' explanations of their choices. The first corpus consists of 1,024 contexts annotated in parallel by three annotators. It captures differences in the identification of coreference across various text types and grammatical-semantic categories, including pronouns, full noun phrases, and anaphoric adverbials. The second corpus comprises 512 contexts, annotated in parallel by five annotators, and focuses on identifying discourse relations in attributive and non-attributive constructions. Both corpora achieve a comparable inter-annotator agreement of approximately 60-65%. For coreference annotation, agreement tends to be lower in cases where automatic coreference resolution models disagree, suggesting that when the models disagree, the examples tend to be more difficult or ambiguous for human annotators to interpret. The annotators' comments, both for coreference and discourse relations, further reveal differences in interpretation, varying levels of confidence in text understanding, and individual reading strategies.

Understanding moral values in social media text offers insight into moral judgement formation, and supervised NLP models trained on crowdsourced data have achieved strong classification performance. However, most approaches simplify the problem by aggregating multiple annotators' labels into a single "ground truth", overlooking the inherent subjectivity of the task. In practice, there are disagreements between annotators caused by personal viewpoint or inherent ambiguities, particularly for short tweets. Here, we extend a pretrained language model with a layer that learns annotator-specific features. Our model improves predictions of individual annotations and yields representations that reveal meaningful insights into annotators' moral perspectives. We show that models trained on aggregated labels may hide variation and give a misleading impression of performance. Overall, we demonstrate that disagreement reflects the inherent subjectivity of the task and that modelling individual perspectives creates benefits for moral classification of texts.

Emotion annotation is inherently subjective, yet most NLP pipelines still assume "gold" labels, typically produced by majority voting, and treat annotator variation as noise. In this paper, we present a multilabel emotion annotation case study and use it to examine how annotator behavior and aggregation choices affect both agreement estimates and downstream emotion classifiers. Rather than collapsing disagreement into a single label, we represent targets as soft vote-share labels (including an intensity-weighted variant) and evaluate models using both thresholded metrics (macro-/micro-F1) and probabilistic alignment (Bernoulli cross-entropy SoftBCE), alongside data-derived disagreement diagnostics. Across annotation regimes, we show that disagreement is structured and leaves measurable traces in model behavior: hard labels may maximize F1 metrics, while soft supervision yields predictions that better reflect empirical annotator variance and uncertainty. Our results provide practical guidance for designing, aggregating, and evaluating multilabel emotion datasets when multiple interpretations are plausible.

Pain is assessed differently by patients, nurses, and clinicians, yet most computational approaches assume a single ground-truth label - effectively ignoring who is doing the rating. We introduce a rater-aware, event-aligned framework that converts sparse, rater-specific pain ratings into discrete pain-change events and aligns continuous wearable physiological signals to these events, preserving rater identity throughout. Applied to multimodal wearable data collected during spine-related pain procedures, the framework identifies substantial disagreement across rater groups and provides preliminary, exploratory evidence of rater-dependent physiological differences preceding reported pain increases. These findings suggest that pain-physiology relationships may not be rater-invariant, and that aggregating assessments across raters may mask meaningful physiological patterns. A rater-aware, event-aligned perspective is therefore a promising direction for interpreting wearable data in real-world clinical pain assessment.

Enthymemes, arguments with unstated premises or conclusions, are pervasive in persuasive discourse, yet their annotation remains notoriously subjective. We present a resource of 1,482 tweets from politically controversial discourse, annotated by five annotators for the presence of enthymemes and their argument structure, designed to study label variation. We first revisit the definition of enthymemes and propose annotation guidelines anchored in Walton's argumentation schemes, offering a structured and constrained approach that nonetheless preserves room for the interpretive nature of the task. This contrasts with past resources, which tend to eliminate disagreement, obscuring its sources and preventing investigation of its potential benefits for model performance. We further propose a complexity analysis of the task, identifying where annotation imposes high cognitive load and may give rise to inconsistent annotation. Our preliminary experiments show that models trained on annotator disagreement outperform models trained on hard majority-vote labels. We close by reflecting on how structural openness in enthymeme definitions and guidelines enables the study of variation in subjective inferential processes for future resources and downstream NLP applications concerned with human inference.

When many people highlight the same document, is the crowd a single consensus, or is it internally structured into reader sub-groups that mark different things -- and is that structure a stable property of a reader or of the document? Building on prior work showing an individual's within-document highlighting signal is a whisper while individuality lives in selection, we ask the group-level question on a co-readership platform using a margin-preserving curveball null. Experiment 1: within a document, readers form strong sub-groups -- pairs agree far beyond what shared salience, mark density, and sentence popularity predict (nearest-neighbour agreement z=+6.3, significant in 88% of documents). Under an eight-block region-preserving null, shared engagement with the same coarse regions of the document accounts for about 40% of this excess; the majority survives as finer reader-specific agreement (z=+3.6, 77% significant). So the within-document crowd is, in a descriptive sense, factional. Experiment 2: is that grouping a stable reader trait? Here we are honest about power. The cross-document split-half reproducibility of a pair's agreement is near zero pooled (+0.078 and 0.000 in two separately drawn samples), and a power calibration shows the test is informative only for pairs that co-read many documents. In the only informative high-overlap subset (k>=4), point estimates are positive but small-sample, imprecise across the separately drawn samples, never significant, and attenuate under the region-preserving null. We therefore leave cross-document stability unresolved: the data is consistent with anything from situational grouping to a weak-to-moderate stable reader trait. The crowd is factional within a document; whether its factions follow the reader across documents is, honestly, beyond our reach.

Current research primarily focuses on model performance, while comparatively less attention has been devoted to uncertainty estimation, particularly in settings where LLMs are increasingly used to generate annotated data. We introduce a framework combining conformal prediction with Collaborative Filtering-style annotators' representation to model LLM behavior in relation to human annotators and to analyze patterns of agreement and disagreement. Using Non-Conformity Scores, we introduce the Ghost Prediction metric and the Ghost Annotator representation to quantify cases in which model predictions diverge from all available human annotations. We compute cosine similarity measures to explore differences in model behavior across sociodemographic axes. We evaluated four LLMs of different size and families across four content moderation datasets. Our finding shows that while we find that all models uncertainty increases with annotator disagreement, larger models tend to be more confident in the classification of texts that are not aligned with any human annotation. Finally, the Ghost Annotator framework reveals a consistent and robust pattern of demographic misalignment, suggesting a structural bias likely rooted in pretraining corpora.

Emotions evolve through the dynamics of conversation, and understanding their transition structure is foundational to applications ranging from mental-health screening to dialogue systems. However, existing studies typically compress multi-rater judgments into a single hard label by majority voting, discarding the uncertainty signal needed to understand turn-to-turn transitions. In this article, we propose Bayesian Spectral Emotion Transition Discovery (BSETD), a two-stage framework that discovers emotion-transition structure from multi-rater soft labels. In the first stage, a hierarchical Dirichlet-Multinomial posterior is constructed through the outer product of soft labels, equipping each cell of the K x K transition matrix with a credible interval and Benjamini-Hochberg (BH) false discovery rate (FDR)-controlled significance. In the second stage, the symmetrized graph Laplacian is spectrally decomposed to separate a low-frequency (inertia) component from a high-frequency (contagion) component. On EmotionLines, BSETD simultaneously recovers the signatures of two distinct affective spaces: the Plutchik-adjacent transitions disgust to anger (log2 lift +0.94) and anger to disgust (+0.86) are over-represented, while the Russell-valence-reversed transitions joy to anger (-0.90) and anger to joy (-0.89) are under-represented. A five-source cross-corpus validation yields pairwise Pearson correlations in 0.91-0.98 within English, 0.79-0.85 against Chinese M3ED, and 0.979 between the human hard labels and the LLM virtual soft labels on the same utterance set, demonstrating that a pipeline preserving annotator uncertainty bridges the computational study of emotion dynamics with established psychological theory.

Human disagreement is ubiquitous and well-known in labeling. However, variation in explanations, captured through token-level human rationales, remains far less explored. At the same time, it is unclear how to best evaluate human labels and rationales -- or even how to best aggregate rationales beyond majority vote -- in light of this variation. Yet, rationales may provide additional insights into the richness of human reasoning, that may differ in style, values and interpretations -- especially in subjective NLP tasks like hate speech detection. In this work, we unify diverse models, training strategies, loss functions, and existing evaluation metrics under a single protocol by systematically re-implementing them across different label and rationale representation spaces. Classification metrics are organized around two key properties -- predictive and distributional -- while explainability metrics through three complementary dimensions: plausibility, faithfulness, and complexity. In this unified supervision framework, we evaluate model behavior across classification and explainability metrics, as well as metric sensitivity to the choice of label (hard and soft) and rationale representation space (hard, intermediate and soft). Results show that both hard and soft metrics favor softer representations, highlighting their effectiveness in capturing variation and the need to rethink evaluation in subjective NLP.

When annotators disagree on a label, the disagreement itself carries signal -- and the number of annotators needed to capture it depends on the evaluation metric. We fine-tune NLI models on label distributions subsampled from ChaosNLI, a dataset providing 100 independent annotator judgments per item, and identify metric-dependent saturation. In our 3-class NLI setting, entropy correlation -- whether the model identifies which items elicit disagreement -- requires N ~ 20-50 annotators to converge, while distributional match (KL divergence) saturates by N ~ 10 (87-95% of improvement across five model seeds). This finding rests on a prior observation: soft labels carry item-specific signal that label smoothing cannot replicate. Across five smoothing intensities, entropy correlation clusters at r ~ 0.45-0.49, while soft labels reach r = 0.643 (p < 0.001); per-item analysis traces this gap to smoothing's inability to distinguish ambiguous items from clear ones. The soft-label advantage replicates across two architectures (DeBERTa, RoBERTa), a non-NLI-pretrained baseline, and an exploratory cross-domain evaluation on content safety. These results suggest that annotation budgets should be informed by the target evaluation metric rather than set uniformly.

Federal agencies are deploying large language models (LLMs) to categorize public comment corpora, where the model's organization of the record shapes what policymakers see and which arguments register. Standard evaluation, anchored on stance accuracy against a small validated set, cannot detect when different models produce materially different categorizations of the same public input. We propose an Interpretive Audit Pipeline that treats multi-model disagreement as diagnostic of interpretive complexity and directs human review toward genuinely ambiguous public input. Analyzing 1,260 public comments on a federal USDA docket across four LLMs, we find that inter-model thematic divergence exceeds within-model prompt variation, and that an expert rubric suppresses deep interpretive disagreement without resolving it. In a two-stage labeling study on a stratified 40-comment subsample, four LLMs and a human annotator labeled independently and then revised after seeing the others' labels. Revision behavior varied across labelers, and the human annotator's revisions frequently introduced framings absent from the ensemble's collective output. We argue disagreement-based evaluation is a necessary complement to accuracy metrics for LLM-assisted interpretive coding.

Annotation quality is difficult to sustain when campaigns span weeks or months with small annotator pools. We present a Setswana sentiment dataset of 3,565 tweets annotated by three native-speaker annotators across eight batches and examine why inter-annotator agreement (IAA) declines over time. Despite an aggregate Randolph's free-marginal Kappa of $κ= 0.76$, "excellent," per-batch $κ$ falls by more than 32 points across the annotation task. Through six targeted analyses, we find that (i) label confusion concentrates on the negative/neutral boundary, (ii) two annotators show run-length drift consistent with autopilot labeling, and (iii) the dominant predictor of $κ$ is temporal simultaneity: tweets labeled within one minute achieve $κ= 0.98$, while those labeled more than a day apart reach only $κ= 0.65$. Annotation speed and tweet-level linguistic features show no meaningful association with $κ$. We benchmark three open multilingual encoders and proprietary models (GPT-5 and Gemini) on three-class sentiment classification; fine-tuning yields gains of 29 to 43 macro-F1 points over pretrained baselines, with GPT-5 few-shot leading overall (62.2 macro-F1). We release the dataset, per-annotation timestamps, and analysis code to support reproducible quality auditing for future African language NLP resources.

Free-form legal essay evaluation in NLP treats expert inter-rater stability as a single ceiling number, and treats LLM-judge agreement with that ceiling as evidence of judge stability. We test both assumptions on the Thai bar examination through an identical-inputs protocol: three Bar Council-trained examiners (A, B, C) and a 26-LLM judge panel score the same 15 cross-graded answers from the same four inputs (question, official Bar Council grading regulation, gold answer, candidate answer). The headline finding is asymmetric. On 10 of 15 cells where the rubric prescribes both axes, all 29 raters converge in a tight band: panel agreement is universal. On the remaining 5 cells where the rubric does not prescribe how to grade a correct final answer that omits a decisive statutory citation, the human panel splits between two coherent readings (B/C majority at the upper rubric band, score 6-8; A minority at the lower band, score 1-2). The LLM judge population does not split symmetrically: 22 of 26 LLMs score in or near B/C's contested band, 3 sit in the regulation-silent middle gap, and only 1 (GPT-5.4 Nano) approaches A's band without consistently scoring within it. Zero LLMs in our 26-judge panel reproduce the minority human reading on the contested cells. The B/C-direction cluster spans every model size, vendor, and price tier we tested. An instrumented three-LLM anchor sub-panel (Claude 4.6 Opus, Gemini 3.1 Pro, GPT-5.4 Pro) carries determinism probes, input ablations, and bootstrap CIs, and reaches anchor panel $α= 0.77$ on the 15 cells against human-panel $α= 0.36$. The high LLM-panel $α$ reflects systematic convergence on the majority reading rather than balanced reproduction of both readings; a benchmark that selects its LLM judge by maximising agreement with a human reference panel will inherit this asymmetry by construction.

Annotator disagreement in emotion classification reflects ambiguity intrinsic to emotion concepts and is essential for predictor-quality assessment in subjective NLP. Yet no prior work integrates soft-label learning with Bayesian deep learning to evaluate uncertainty along axes including annotator-distribution fidelity. We train a linear head on a frozen RoBERTa via cyclical stochastic gradient Markov chain Monte Carlo (cSG-MCMC), targeting the empirical annotator distribution with a soft-label objective under a five-axis evaluation. On the 28-emotion GoEmotions benchmark, the proposed method outperforms Monte Carlo Dropout and Deep Ensemble simultaneously on three axes -- Jensen-Shannon divergence (JSD) to the annotator distribution, Spearman correlation between per-emotion aleatoric uncertainty and disagreement, and selective-prediction Area Under the Risk-Coverage Curve (AURC) and Area Under the ROC Curve (AUROC) -- showing independent axes are jointly attainable from one posterior. Post-hoc temperature scaling exhibits a bidirectional effect, establishing hard-label calibration and annotator-JSD as independent dimensions and motivating joint reporting as an honest protocol.