When training machine learning models, we typically treat our datasets as collections of independent and identically distributed (i.e., i.i.d.) observations. We assume that when a human annotator evaluates a data point—whether it is a tweet, an audio clip, or a sentence—their judgment depends entirely on the content of that specific item.
However, humans are not memoryless machines. They carry the cognitive residue of what they have just seen into what they evaluate next. This phenomenon, where the presentation order of data systematically skews human judgment, is known as sequential annotation bias. Because standard aggregation methods (like majority voting) assume independent errors, these systematic psychological quirks bake hidden, non-random dependencies directly into our training data.
The Cognitive Mechanisms of Order Effects
Why does the sequence of data alter an annotator's judgment? Several distinct psychological mechanisms explain how the immediate past skews the present.
Anchoring and Priming
The most immediate form of sequential bias is anchoring, where an annotator uses their previous score as a reference baseline for the current item. Mathur et al. demonstrated this in continuous-scale labeling tasks, showing a strong positive correlation between consecutive labels. They also observed the gambler's fallacy: if an annotator assigns the same label several times in a row, they become artificially prone to changing their next label simply because they expect a shift in distribution.
This bias extends to emotional and contextual priming. Mieleszczenko-Kowszewicz et al. found that the emotional weight of previously processed text bleeds into subsequent judgments. For instance, exposure to a sequence of highly negative texts primes an annotator to perceive a following neutral or ambiguous text more negatively than they would have in an isolated context.
Contextual vs. Isolated Presentation
The structural presentation of data dictates how these biases manifest. Jaiswal evaluated this by comparing emotion labels generated from continuous conversational streams against those generated from randomized, isolated clips. The author found that temporal context fundamentally alters human perception; labels from continuous streams capture conversational dynamics, whereas randomized clips strip away the sequential context, leading to highly inconsistent ground-truth data between the two methods.
Independent Assumption (Standard ML):
[Item A] -> Label A
[Item B] -> Label B
[Item C] -> Label C
Sequential Reality (Human Cognitive Flow):
[Item A] -> Label A
│ (Priming / Anchoring)
▼
[Item B] -> Label B'
│ (Fatigue / Satisficing)
▼
[Item C] -> Label C'
Interface Design and Cognitive Fatigue
Sequential bias is not just driven by the data itself; it is heavily modulated by the design of the annotation interface and the length of the task.
- Primacy Effects: The order of operations within an interface creates systematic imbalances. Kern et al. introduced the concept of "annotation sensitivity," demonstrating that the sequence and grouping of classification categories alter the final label distribution. Specifically, categories presented earlier to an annotator experience a primacy effect, receiving a higher frequency of labels simply because they are encountered first. Similarly, Fournier et al. noted that a fixed evaluation order across different emotional dimensions introduces systematic dependencies, where ratings on initial dimensions bias the scores of subsequent ones.
- Satisficing: As annotators progress through a long queue of items, cognitive fatigue sets in, leading to a phenomenon known as satisficing—where workers reduce their cognitive effort to complete the task faster. Beck et al. discovered that later items in an annotation sequence are significantly less likely to be assigned heavy or sensitive labels, such as hate speech or offensive language.
- Structural Heuristics: When tasks are complex, annotators often adopt lazy structural shortcuts. Lotfi et al. showed that in knowledge-grounded dialogue tasks, crowd-workers frequently "follow the knowledge" by selecting information based on its sequential presentation order within a provided knowledge pool, creating structural artifacts that do not reflect genuine human knowledge flow.
Modeling Annotators as Sequential Agents
If human annotations are inherently sequential, how can we recover the true "ground truth" labels? The solution is to move away from independent token aggregation and instead model annotators as sequential decision-makers.
Sequential Confusion Matrices
In sequence labeling tasks like Named Entity Recognition (NER), annotators make structured, dependent errors—such as missing the start of a text span but correctly identifying subsequent tokens once they realize they are inside the span. To capture this, Simpson et al. developed a Bayesian framework that utilizes a sequential confusion matrix. Instead of assuming an annotator's error rate is constant, their model conditions the annotator's current label on both the true underlying label and the annotator's own previous label.
$$\mathbb{P}(\text{Label}_t \mid \text{True}_t, \text{Label}_{t-1})$$
By formalizing this transition dependency, the model can mathematically infer when an annotator has prematurely terminated a sequence or missed a transition boundary.
Transition-Based Dependencies
Similarly, Lu et al. treated crowd workers as sequential agents, modeling both local (neighboring words) and non-local (repeated words) transitions. This allows aggregation algorithms to separate general annotator confusion from structured transition errors. Huang et al. also demonstrated that incorporating these sequential dependencies into probabilistic crowd-labeling models dramatically improves ground-truth estimation for tasks like discourse segmentation, where consecutive labels are inherently linked.
Turning Bias into a Feature
Rather than merely mitigating sequential bias, some frameworks actively exploit it. Because human beings are naturally better at making relative comparisons than absolute judgments, Naini et al. leveraged the anchoring effect on consecutive annotations to derive highly reliable preference and ordinal labels. By analyzing how an annotator's judgment shifts from one sample to the next, the authors transformed what is normally considered data noise into a structured signal for preference learning.
Key Foundational Work
Sequence Effects in Crowdsourced Annotations (2017) — contributed 4 concept elements to this direction.
Estimation of Discourse Segmentation Labels from Crowd Data (2015) — contributed 1 concept elements to this direction.