The Label Bias Problem in Sequence Models

54 concepts38 papers1 proposal

When building models that predict sequences—whether parsing a sentence, transcribing speech, or generating text—we often design them to make decisions step-by-step. At each discrete step, the model looks at its current state and outputs a probability distribution over the next possible actions or labels.

To make this a valid probability distribution, the scores for all outgoing transitions from the current state must sum to one. This localized accounting system is known as local normalization. While computationally efficient, local normalization introduces a systemic mathematical flaw: the label bias problem.

The Conservation of Probability Mass

To understand why label bias occurs, imagine a model navigating a state space structured like a maze. At state $A$, the maze splits into one hundred different paths. At state $B$, there is only a single, narrow corridor leading forward.

If the model is locally normalized, it must distribute a fixed probability budget of exactly $1.0$ among the outgoing paths at every single step. At state $A$, that budget is split one hundred ways, meaning each individual path receives a tiny fraction of the probability mass. At state $B$, however, the single outgoing path receives the entire budget of $1.0$ by default.

State A (High Branching):  P(path_1) + P(path_2) + ...

+ P(path_100) = 1.0  -> Each path gets ~0.01
State B (Low Branching):   P(path_1) = 1.0                                 -> Path gets 1.0 automatically

What happens if the actual input data (the observation) strongly suggests that the path out of state $B$ is incorrect? The model has no choice but to ignore the evidence. Because the local transitions must sum to one, and there are no alternative paths to absorb the probability mass, the model assigns a probability of $1.0$ to the transition anyway.

As Mark Johnson 0001 demonstrated in Joint and Conditional Estimation of Tagging and Parsing Models, this structural constraint forces conditional models to ignore observation data in favor of the state sequence topology. Similarly, Phil Blunsom observed in Maximum Entropy Markov Models for Semantic Role Labelling that directed discriminative models will effectively ignore input features when trapped in states with low branching factors. The model becomes a victim of its own graph topology, as noted by Brian D Ziebart et al. in Maximum Entropy Inverse Reinforcement Learning.

Label Bias vs. Observation Bias

It is helpful to distinguish label bias from a related but distinct pathology: observation bias. In their seminal paper, Conditional Structure versus Conditional Estimation in NLP Models, Christopher D Manning et al. disentangled these two phenomena:

  • Label Bias is driven by the absence of competition. Because a state has few outgoing transitions, the model is forced to follow them regardless of the observation. The structure overrides the input.
  • Observation Bias is driven by the strength of a local observation. It occurs when a highly informative local feature effectively "explains away" and overrides the preceding sequence context. The immediate input overrides the history.

Both biases stem from local normalization, but they attack the model's contextual awareness from opposite directions.

How Label Bias Manifests in Modern Deep Learning

Label bias is not confined to classic models like Maximum Entropy Markov Models (MEMMs); it actively degrades modern deep learning architectures across several domains.

Transition-Based Parsing

In incremental, transition-based natural language processing, a model predicts a sequence of actions (such as shift or reduce) to construct a parse tree. Hao Zhou 0012 et al. in A Neural Probabilistic Structured-Prediction Method for Transition-Based Natural Language Processing showed that locally normalized models suffer when forced to assign a high probability to a mediocre action simply because all other local alternatives are even less likely. Once an early error is made, the model becomes trapped in a low-probability region of the search space and cannot use future context to self-correct, a limitation highlighted by Daniel Andor et al. in Globally Normalized Transition-Based Neural Networks.

Sequence-to-Sequence Generation

In autoregressive text generation and neural machine translation, local normalization penalizes states with many outgoing paths. As Sam Wiseman et al. explained in Sequence-to-Sequence Learning as Beam-Search Optimization, a model might unfairly penalize a transition into a rich, expressive vocabulary space because the probability mass must be distributed widely, even if that path eventually yields a higher-quality global sequence.

Furthermore, David Chiang et al. in Correcting Length Bias in Neural Machine Translation identified that the end-of-sentence (<EOS>) token often creates a zero-entropy suffix distribution. Because the model cannot look ahead to evaluate global sequence quality, it overestimates the probability of terminating early, leading to a systemic short-sentence bias.

Imitation Learning and Inverse Reinforcement Learning

When training agents to match expert trajectories, objectives that focus purely on matching local decisions suffer from label bias. As Gergely Neu et al. argued in Training parsers by inverse reinforcement learning, an agent might achieve high local accuracy but fail completely to recover the correct global trajectory because local optimization fails to respect the global constraints of the structured output space.

The Solution: Global Normalization

To eliminate label bias, we must remove the requirement that the model distribute a fixed probability budget at each individual step. Instead of local normalization, we use global normalization.

In a globally normalized model, the system assigns unnormalized scores (often called energies or potentials) to transitions at each step. The total score of an entire sequence is simply the sum of its step-level scores. Normalization happens exactly once—at the very end—across all possible complete sequences:

$$P(Y|X) = \frac{\exp(\sum_t \text{score}(y_t, y_{t-1}, X))}{Z(X)}$$

where $Z(X)$ is the global partition function summing over every valid path through the sequence space.

Locally Normalized:   Step 1 (Sum=1) -> Step 2 (Sum=1) -> Step 3 (Sum=1)
Globally Normalized:  Step 1 (Score)  -> Step 2 (Score)  -> Step 3 (Score) -> Normalize Entire Path

This subtle shift completely alters the model's behavior:

  1. Retroactive Correction: If a path looks promising early on but leads to a dead end or a highly incoherent state later, a massive negative score at step 10 can retroactively penalize the entire sequence.
  2. Freedom from Topology: A state with only one outgoing transition is no longer forced to give it a probability of $1.0$. If the transition is bad, it receives a low score, lowering the probability of the entire path.

Michael Collins bypassed local limitations by introducing the Structured Perceptron in Discriminative Training Methods for Hidden Markov Models: Theory and Experiments with Perceptron Algorithms, allowing models to optimize for global sequence features. Conditional Random Fields (CRFs), as detailed by Charles Sutton et al. in An Introduction to Conditional Random Fields, serve as the standard probabilistic architecture for achieving this global balance.

Modern deep learning continues to leverage this insight. Kai Hu et al. integrated this into Neural CRF transducers for sequence labeling to retain long-range dependencies without local traps. Similarly, Milos Stanojevic et al. demonstrated in Max-Margin Incremental CCG Parsing that a global max-margin objective allows incremental parsers to evaluate competing hypotheses across an entire search beam simultaneously, completely neutralizing the local normalization tax.

Key Foundational Work

Conditional Structure versus Conditional Estimation in NLP Models (2002) — contributed 5 concept elements to this direction.
Joint and Conditional Estimation of Tagging and Parsing Models (2001) — contributed 4 concept elements to this direction.
Ranking Algorithms for Named Entity Extraction: Boosting and the Voted Perceptron (2002) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 41 papers
  2. Version 1 · 2026-10-04 · 56 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_label_bias_bb016278,
  title = {label bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/label-bias/rdr_bb016278005746979b4d0a6b77742058},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_label_bias_bb016278_v2,
  title = {label bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/label-bias/rdr_bb016278005746979b4d0a6b77742058/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake