When a machine learning model fails to generalize to a new dataset, we typically diagnose the failure as "domain shift." We assume the model is tripping over unfamiliar vocabulary, novel syntactic structures, or shifts in the underlying topic distribution. However, a substantial body of research reveals a more insidious and frequently overlooked culprit: annotation divergence.
Annotation divergence occurs when two datasets represent the exact same underlying reality but use different, often arbitrary, rules to label it. As Guillaume Wisniewski et al. demonstrate in How Bad are PoS Tagger in Cross-Corpora Settings? Evaluating Annotation Divergence in the UD Project, when you align the annotation style of a test set to match the training set's conventions, the perceived "domain gap" largely evaporates, bringing performance back to near in-domain levels. This implies that models frequently do not fail because they misunderstand the language or data; they fail because they are being evaluated against a moving ground-truth target.
The Anatomy of Labeling Inconsistencies
To build intuition for why this happens, consider how human experts create datasets. Every annotation effort relies on a set of guidelines to resolve ambiguous edge cases. Different teams inevitably make different choices, creating systematic mismatches across several dimensions:
- Span Boundaries: In sequence labeling tasks like Named Entity Recognition (NER) or temporal extraction, does an entity span include preceding determiners or trailing prepositions? Rodrigo Agerri et al. in Robust Multilingual Named Entity Recognition with Shallow Semi-Supervised Features and Steven Bethard et al. in Extracting Time Expressions from Clinical Text note that sequence labelers suffer severe penalties when moving between corpora simply because one dataset includes a preposition inside a time expression while another excludes it.
- Structural Definitions: In dependency parsing, how should a specific grammatical marker or prepositional phrase attach to a verb? Joakim Nivre et al. observe in The CoNLL 2007 Shared Task on Dependency Parsing that these structural schema discrepancies degrade parsing performance far more than differences in vocabulary.
- Visual Boundaries: This issue is not unique to natural language processing. In document analysis, Khurram Azeem Hashmi et al. highlight in Guided Table Structure Recognition through Anchor Optimization how table structure recognition models struggle because one dataset marks only text regions while another marks full-width structural cells. Similarly, in 3D object detection, Dusan Malic et al. point out in SAILOR: Scaling Anchors via Insights into Latent Object Representation that datasets differ on whether a vehicle's bounding box should encompass its side mirrors.
Why Statistical Adaptation Hits a Wall
Why can't our best unsupervised domain adaptation techniques fix this?
Traditional unsupervised adaptation methods—such as aligning feature spaces or performing masked language modeling on target text—operate entirely on the input distribution $P(X)$. They attempt to make the representation of the target domain look like the source domain.
But as Jacob Eisenstein argues in Part-of-Speech Tagging for Historical English and the authors further elaborate in Unsupervised Domain Adaptation of Contextualized Embeddings for Sequence Labeling, annotation drift is a shift in the conditional distribution $P(Y|X)$ driven by human choices, not natural data evolution. If two datasets label the exact same phrase differently based on conflicting guidelines, no amount of unsupervised feature alignment can reconcile the contradiction. The model learns to distinguish between the underlying annotation styles rather than the true features, creating an artificial performance ceiling (Anders Søgaard in Data point selection for cross-language adaptation of dependency parsers; Joachim Wagner 0001 et al. in Treebank Embedding Vectors for Out-of-domain Dependency Parsing).
The Severity of the Schema Gap
Remarkably, the impact of divergent annotation styles often dwarfs the impact of actual subject-matter shifts. Barbara Plank et al. isolated this effect in relation extraction by comparing models across scientific sub-domains in What Do You Mean by Relation Extraction? A Survey on Datasets and Study on Scientific Relation Classification.
The authors discovered that models experience sharper performance declines when moving between different labeling conventions within the same field than when moving between entirely different sub-topics that share a consistent annotation style. The schema mismatch introduces a systematic error that purely data-driven models cannot natively bypass.
Modern Implications and the LLM Dilemma
As the field shifts toward leveraging Large Language Models (LLMs) for zero-shot data annotation or model adaptation, this problem takes on a new form. One might hope that an LLM could act as a universal annotator to bypass human inconsistencies.
However, Seiji Shimizu et al. demonstrate in Exploring LLM Annotation for Adaptation of Clinical Information Extraction Models under Data-sharing Restrictions that directly using LLMs to generate labels for smaller, specialized models often causes performance declines. The reason is identical: the LLM's intrinsic zero-shot labeling style frequently deviates from the specific formatting, stylistic constraints, and conceptual boundaries of the original training data. To build truly robust systems, we must treat the annotation schema not as a static truth, but as a dynamic variable that requires explicit alignment.
Key Foundational Work
Frustratingly Hard Domain Adaptation for Dependency Parsing (2007) — contributed 8 concept elements to this direction.
The CoNLL 2007 Shared Task on Dependency Parsing (2007) — contributed 3 concept elements to this direction.
Broad-Coverage Sense Disambiguation and Information Extraction with a Supersense Sequence Tagger (2006) — contributed 1 concept elements to this direction.