The Gold Standard of Judgement | Conditional Random Fields, Feature Functions, Structured Prediction and Sequence Labels

Many decisions arrive as structures rather than isolated labels. Words form sequences, image pixels touch neighbours, biological positions have order and document fields follow layouts. Classifying each element independently can create impossible or inconsistent outputs even when the local scores look plausible.

Conditional random fields, or CRFs, model a conditional distribution over structured labels given observed input. A linear-chain CRF can combine input features with transition features and choose a coherent sequence. This guide extends the Markov-random-field article: it focuses on discriminative conditional modelling, feature functions, sequence inference, training and evidence checks rather than a general undirected joint model.

Choose a reading route

Open the full chapter index · Use the field check · Return to the Judgement root

Chapter contents

Chapters 1–3 · Conditional structure and features

1. 1. Model the conditional distribution the task actually needs
2. 2. Use a linear chain when neighbouring labels carry the relevant structure
3. 3. Define feature functions that expose evidence and assumptions

Chapters 4–6 · Scoring, inference and worked sequence

1. 4. Understand log-linear sequence scores and global normalisation
2. 5. Separate marginals, Viterbi decoding and constrained decisions
3. 6. Worked example: label a fictional library request

Chapters 7–9 · Training, constraints and evaluation

1. 7. Train weights with conditional likelihood and expected features
2. 8. Use regularisation and feature selection to control variance
3. 9. Encode label constraints without hiding annotation disagreements

Chapters 10–12 · Transfer, governance and audit

1. 10. Evaluate complete structures, uncertainty and error cost
2. 11. Combine modern encoders with CRFs without losing the evidence boundary
3. 12. Audit a CRF from annotation to deployed decision

A CRF models p(y|x): the distribution of an output structure y given observed input x. It does not need to model how the input itself was generated. This discriminative focus permits overlapping and correlated input features while retaining dependencies among output labels.

Define the observation structure, label structure and prediction target before choosing a graph. State whether the task needs per-position marginals, the most likely complete labeling or several plausible structures. Compare with independent classification and a generative model so the value of structured dependence is testable.

For a sentence, x contains tokens and observed features while y contains entity tags. The CRF scores entire tag sequences conditioned on the sentence instead of assigning each token a label in isolation.

Changed-context transfer check

Transfer the definition to an image grid or a biological sequence. Identify what is observed, what is predicted and which output dependencies are substantively plausible.

Limit and misconception

Conditioning avoids a model of p(x), but it does not remove assumptions about graph structure, features, labels, training population or missing data.

Evidence practice

Keep the observation definition, label space, graph, prediction target and baseline models. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: using a CRF because data are sequential without stating which output dependencies matter. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Next chapter

A linear-chain CRF connects labels at adjacent positions and associates them with the observed sequence. It is appropriate when local transitions capture the main output structure and dynamic programming remains tractable. More complex dependencies require a different graph or approximation.

Draw one label variable per position and define factors for transitions and state-observation compatibility. State the order and boundary conditions. Test whether long-range dependencies, segment lengths or nested labels are essential; if so, compare a semi-Markov, higher-order or other structured model rather than hiding the mismatch.

In a name-tagging task, an inside-person label may follow a begin-person label, while an isolated inside label is discouraged. The chain represents that local grammar without claiming words cause one another.

Changed-context transfer check

Change the task to nested entities or events spanning distant clauses. Explain which relationships a first-order chain cannot represent directly.

Limit and misconception

A chain is a computational and modelling choice, not a discovery that reality has only nearest-neighbour dependence.

Evidence practice

Keep the chain graph, factor scopes, order, boundaries and tests for omitted dependence. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: forcing a complex structure into first-order transitions and interpreting errors as random noise. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Next chapter

Feature functions connect observations and candidate labels to numeric contributions. State features can use token, shape, context or encoder evidence; transition features describe adjacent label combinations. Their weights are learned, but feature design determines what evidence the model can use.

Name each feature, scope, transformation, missing-value rule and intended interpretation. Separate features available at prediction time from leaked future or label-derived information. Inspect activation frequency, rare conjunctions, correlation and subgroup coverage before training.

A capitalisation feature supports a person tag, a lexicon feature supports an organisation tag and a transition feature discourages I-PER after O. None is decisive alone; the sequence score combines them.

Feature type	Example job	Failure to test
State	Relate input to current label	Template leakage
Transition	Relate adjacent labels	Impossible sequence still allowed
Window	Use nearby observations	Boundary handling
Encoder	Provide learned representation	Domain and subgroup drift

Changed-context transfer check

Transfer the feature audit to a document-layout model. Identify features that might encode scanner, template or institution rather than the intended field.

Limit and misconception

A readable feature is not necessarily causal, fair or stable. High-dimensional learned encoders can hide the same risks rather than eliminate them.

Evidence practice

Keep the feature dictionary, scopes, data availability, activation counts and leakage review. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: including information unavailable at inference or treating proxy features as neutral evidence. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Next chapter

A linear-chain CRF sums weighted feature values across a candidate label sequence, exponentiates that score and divides by a partition function over all possible sequences for the same input. Global normalisation lets competing complete labelings be compared consistently.

Write the score, feature sum and input-dependent partition function explicitly. Use numerically stable log-sum-exp calculations and test a tiny sequence by exhaustive enumeration. Verify that probabilities over all sequences sum to one and that adding a constant to all scores does not change them.

For a two-token sequence with two labels, four candidate sequences can be enumerated. A strong local score for one label may lose to another sequence because transition evidence makes the full structure more coherent.

Changed-context transfer check

Change one transition weight and calculate which sequences gain probability. Explain why a large weight can overwhelm useful local evidence.

Limit and misconception

A globally normalised probability is conditional on the chosen model and parameters. It is not automatically calibrated or a probability that the real-world assertion is true.

Evidence practice

Keep the feature values, weights, sequence scores, partition function and enumeration benchmark. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: treating unnormalised scores as probabilities or ignoring numerical overflow and underflow. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Next chapter

Forward-backward computes label or transition marginals in a linear chain, while Viterbi decoding finds the highest-scoring complete sequence. Marginally most likely labels need not form the highest-probability sequence, and a downstream decision may require costs or constraints beyond either output.

Declare the inference target. Compare forward-backward marginals with exhaustive enumeration on short sequences and Viterbi output with the maximum enumerated score. Preserve ties, near-ties and impossible labels. If a business rule changes the chosen output, implement and audit it as a separate decision layer.

The Viterbi sequence labels an entire phrase as an organisation, while one token has only 0.54 marginal support. The report shows both the coherent sequence and the local uncertainty for review.

Changed-context transfer check

Transfer the distinction to a medical-text extractor where missing one critical field costs more than adding an extra candidate. Design a review threshold without rewriting the model probability.

Limit and misconception

Exact chain inference does not validate features, training data or label meaning. A precisely decoded wrong model remains wrong.

Evidence practice

Keep the declared target, forward-backward tables, Viterbi backpointers, ties and decision rules. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: reporting a best sequence as certainty or mixing marginal labels into an incoherent sequence. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Next chapter

A fictional help desk extracts book titles, dates and branches from short requests. One message reads “Reserve River Light for Monday at North Branch”. The label scheme uses BIO tags so multi-token fields remain contiguous.

Local features support “River” and “Light” as title tokens, “Monday” as a date and “North Branch” as a location. Transition features favour B-TITLE followed by I-TITLE and reject I-LOCATION after O. Viterbi produces one coherent sequence; marginals reveal uncertainty over whether “North” begins a location or modifies another field.

A librarian reviews low-margin spans and corrects two annotation inconsistencies. The team updates guidelines before retraining and evaluates on requests from different branches rather than only the template used for development.

Token	Candidate evidence	Review issue
River	Title lexicon and context	Could begin a person or title
Light	Title continuation	Transition supports I-TITLE
Monday	Date pattern	Possessive form may differ
North Branch	Location phrase	Boundary and branch vocabulary

Changed-context transfer check

Change the request to “Reserve Monday’s River Light edition”. Explain how possessive form and ambiguous title boundaries challenge the same features.

Limit and misconception

The example is fictional. Automated extraction should support, not silently replace, accountable review when errors affect access, rights or important services.

Evidence practice

Keep the annotated sequence, feature activations, scores, marginals, decoded path and review corrections. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: celebrating one correct path while annotation rules and ambiguous spans remain unresolved. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Next chapter

CRF training commonly maximises regularised conditional log-likelihood. The gradient compares feature counts in labelled data with their expectations under the current model. Forward-backward supplies those expectations for a linear chain, while optimisation adjusts weights across many sequences.

Fix train, validation and test splits before tuning. Reproduce feature counts on tiny data, check analytic gradients against finite differences and monitor objective and gradient norm. Record optimiser, learning schedule, stopping rule, random seed and numerical precision.

A transition appears often in training but mostly within one document template. Its weight grows until validation on another source worsens. Stratified evaluation reveals that frequency alone did not establish general usefulness.

Changed-context transfer check

Transfer the learning check to partially labelled sequences. State which objective or marginalisation handles unknown labels instead of treating them as negative examples.

Limit and misconception

Optimising likelihood can improve fit without improving the decision metric, calibration or fairness relevant to use.

Evidence practice

Keep the labelled splits, feature counts, expected counts, gradient checks, optimiser state and learning curves. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: evaluating on training sequences or trusting an optimiser that cannot reproduce a tiny benchmark. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Next chapter

CRFs can contain many sparse, overlapping features. L2 regularisation shrinks weights smoothly, while L1 can encourage sparsity; both encode assumptions about useful complexity. Feature templates and learned encoders still require enough representative evidence.

Tune regularisation on a validation set that matches deployment conditions. Compare simple baselines and ablate major feature groups. Inspect weight distributions, rare high-weight features and performance across sources. Preserve the full tuning path rather than only the winning value.

A rare header string receives a large weight and boosts development accuracy but fails on another library system. Stronger regularisation and template ablation expose the shortcut.

Changed-context transfer check

Apply the test to a neural encoder plus CRF. Decide which capacity is controlled by encoder training and which by the transition layer.

Limit and misconception

Sparsity does not prove interpretability or truth. Regularisation can also suppress rare but important patterns for underrepresented groups.

Evidence practice

Keep the regularisation grid, validation results, ablations, weight review and subgroup errors. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: selecting complexity from the test set or keeping brittle rare features because aggregate accuracy rises. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Next chapter

BIO, BILOU and other schemes translate spans into sequence labels. Transition constraints can prevent structurally invalid paths, but they cannot repair inconsistent definitions of what counts as an entity, event or boundary. Annotation policy is part of the model.

Publish label definitions with positive, negative and edge examples. Measure annotator agreement at token and span level, adjudicate disagreements and version guidelines. Implement allowed transitions as explicit masks and test every label pair, including sequence start and end.

Annotators disagree whether honorifics belong inside person spans. A hard transition rule cannot resolve the semantic question, so the team updates the guideline and relabels affected training cases.

Changed-context transfer check

Transfer the policy to nested or overlapping spans. Explain why a single flat BIO chain cannot represent every valid structure.

Limit and misconception

Constraints can make outputs look clean while forcing uncertain or novel cases into the nearest allowed label. Preserve an unknown or review path where appropriate.

Evidence practice

Keep the label manual, annotated examples, agreement results, adjudications, transition mask and version history. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: using hard constraints to conceal inconsistent training labels or unsupported ontology choices. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Next chapter

Token accuracy can be high while complete spans are wrong, especially when the outside label dominates. Evaluation should match the structured task and downstream harm: span precision, recall and F1, exact sequence accuracy, calibration, subgroup performance and review workload may all matter.

Keep a locked test set, report per-label and span metrics, inspect boundary and transition errors and calculate confidence reliability. Evaluate shifted sources and ambiguous cases. Choose review thresholds using explicit false-positive and false-negative costs rather than one convenient aggregate.

A model reaches 97 percent token accuracy but misses many short date spans because most tokens are O. Span recall and branch-specific results reveal the failure hidden by the aggregate.

Measure	What it asks	Blind spot
Token accuracy	How many positions match?	Dominant outside label
Span F1	Are complete fields recovered?	Unequal error cost
Sequence accuracy	Is the whole structure correct?	Harsh on long sequences
Calibration	Does confidence match frequency?	Can drift after deployment

Changed-context transfer check

Change the case so false positives trigger unnecessary manual work but false negatives deny service. Design a threshold and escalation rule that reflects both consequences.

Limit and misconception

Metrics estimated from historic labels inherit their omissions and disagreements. Calibration can fail after drift even when ranking remains useful.

Evidence practice

Keep the locked test set, token and span metrics, calibration, confusion patterns, subgroup results and cost rule. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: publishing token accuracy alone or choosing a threshold without the people and decisions affected. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Next chapter

A neural encoder can supply contextual emissions while a CRF layer models output transitions. End-to-end training may improve performance, but the CRF does not make the encoder interpretable, remove data bias or guarantee that transition structure adds value.

Compare encoder-only and encoder-plus-CRF systems under the same splits. Freeze versions, tokenisation, truncation, pretrained weights and transition constraints. Test ablations, out-of-domain data and long or unusual sequences. Inspect whether gains come from valid structure or benchmark-specific formatting.

A CRF improves span consistency on one corpus but adds no benefit after a stronger encoder on another. The team retains it only where replicated gains and review needs justify the complexity.

Changed-context transfer check

Transfer the comparison to multilingual text with different tokenisation and entity conventions. Identify evidence needed before sharing one transition matrix.

Limit and misconception

A more coherent output can still be confidently wrong. Language models and CRFs can propagate annotation and social biases into material decisions.

Evidence practice

Keep the encoder and tokenizer versions, transition layer, ablations, shift tests and error review. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: assuming a CRF layer automatically improves every structured task or makes a neural model safe. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Next chapter

A reproducible audit follows label policy, source data, feature or encoder construction, graph, objective, regularisation, inference, metrics, thresholds and human review. It distinguishes model probability from decision authority and records corrections after deployment.

Version data and code, reproduce training from seeds, run tiny enumeration and gradient tests, validate transition masks and compare saved predictions with recomputed outputs. Trace sampled decisions to input, annotations, scores, marginals, decoded path and reviewer action. Monitor drift and appeal outcomes.

A fictional audit finds that a tokenizer update shifted span boundaries while aggregate F1 barely moved. Reviewing corrected cases exposes repeated location errors, so the team restores the prior tokenizer and revalidates.

Changed-context transfer check

Design the audit for privacy-protected text that cannot be retained. Specify approved snapshots, hashes, feature summaries and secure replay procedures.

Limit and misconception

Technical reproducibility does not by itself establish lawful, fair or appropriate use. Governance and remedy must match the consequence of errors.

Evidence practice

Keep the label policy, data lineage, code, model artefact, tests, predictions, review actions and correction log. Label every field as observed, computed, assumed or chosen. Give the record to a second reader and ask them to reproduce the result without the original author explaining what was meant. Differences reveal hidden definitions, missing versions or decision rules that still need to be made explicit.

Deliberately test for this failure: releasing a model score while no reviewer can reconstruct the label path or challenge its use. Record the earliest point at which the process can detect it, the correction applied and whether later conclusions change. A method is transferable only when its controls work on an altered case, not merely on the example used to teach it.

Contents · Previous chapter · Connected articles

Advanced boundary cases

Semi-Markov and segmental CRFs

When spans rather than adjacent labels are the natural units, segmental models can represent duration and span features directly. Their added computation is justified only when validation shows that the simpler chain is inadequate.

Partial and weak supervision

Missing or rule-generated labels require objectives that preserve uncertainty. Do not silently convert unknown positions into outside labels or report weak labels as ground truth.

Higher-order and skip dependencies

Longer-range label relationships can improve fit but enlarge inference and overfitting risk. State the structural hypothesis, benchmark exactness or approximation and retain a simpler baseline.

A practical field check

Use this check before training, evaluating or deploying a CRF or another structured sequence predictor.

1. Define observations, labels and the structured prediction target.
2. Justify the graph and factor scopes.
3. Version label rules and feature or encoder inputs.
4. Benchmark normalisation, marginals and decoding on tiny cases.
5. Validate gradients, regularisation and stopping rules.
6. Evaluate spans, sequences, calibration and subgroup errors.
7. Separate model output from decision thresholds and review.
8. Preserve drift, correction and appeal evidence.

Run the check as a small workbook

Use sheets for label definitions, sequences, features, transition rules, training runs, predictions, metrics, calibration, review and corrections. Every predicted span should trace to the model version and annotation policy that produced it.

Gate	Pass condition	Stop condition
Labels	Definitions and boundaries are versioned	Annotation disagreement hidden
Model	Features, graph and normalisation reproduce	Scores treated as probabilities
Evaluation	Structured metrics and shift tests pass	Token accuracy used alone
Use	Thresholds, review and correction are governed	Prediction becomes unappealable fact

Chapter-by-chapter evidence audit

Chapter	Evidence to keep	Failure to catch
1. Model the conditional distribution the task actually needs	The observation definition, label space, graph, prediction target and baseline models	Using a CRF because data are sequential without stating which output dependencies matter
2. Use a linear chain when neighbouring labels carry the relevant structure	The chain graph, factor scopes, order, boundaries and tests for omitted dependence	Forcing a complex structure into first-order transitions and interpreting errors as random noise
3. Define feature functions that expose evidence and assumptions	The feature dictionary, scopes, data availability, activation counts and leakage review	Including information unavailable at inference or treating proxy features as neutral evidence
4. Understand log-linear sequence scores and global normalisation	The feature values, weights, sequence scores, partition function and enumeration benchmark	Treating unnormalised scores as probabilities or ignoring numerical overflow and underflow
5. Separate marginals, Viterbi decoding and constrained decisions	The declared target, forward-backward tables, Viterbi backpointers, ties and decision rules	Reporting a best sequence as certainty or mixing marginal labels into an incoherent sequence
6. Worked example: label a fictional library request	The annotated sequence, feature activations, scores, marginals, decoded path and review corrections	Celebrating one correct path while annotation rules and ambiguous spans remain unresolved
7. Train weights with conditional likelihood and expected features	The labelled splits, feature counts, expected counts, gradient checks, optimiser state and learning curves	Evaluating on training sequences or trusting an optimiser that cannot reproduce a tiny benchmark
8. Use regularisation and feature selection to control variance	The regularisation grid, validation results, ablations, weight review and subgroup errors	Selecting complexity from the test set or keeping brittle rare features because aggregate accuracy rises
9. Encode label constraints without hiding annotation disagreements	The label manual, annotated examples, agreement results, adjudications, transition mask and version history	Using hard constraints to conceal inconsistent training labels or unsupported ontology choices
10. Evaluate complete structures, uncertainty and error cost	The locked test set, token and span metrics, calibration, confusion patterns, subgroup results and cost rule	Publishing token accuracy alone or choosing a threshold without the people and decisions affected
11. Combine modern encoders with CRFs without losing the evidence boundary	The encoder and tokenizer versions, transition layer, ablations, shift tests and error review	Assuming a CRF layer automatically improves every structured task or makes a neural model safe
12. Audit a CRF from annotation to deployed decision	The label policy, data lineage, code, model artefact, tests, predictions, review actions and correction log	Releasing a model score while no reviewer can reconstruct the label path or challenge its use

Connected articles across the eduKate ecosystem

Start with the governing route

• The Gold Standard of Judgement
• The Gold Standard of Education | Learning That Transfers into Life
• How X Works hub

Continue into verified specialist studies

• Markov Random Fields, Factor Graphs, Undirected Dependence and Message Passing — place the conditional model inside the wider undirected graphical-model family.
• Bayesian Networks, Conditional Independence, Diagnostic Inference and Sensitivity — contrast directed generative structure with conditional structured prediction.
• Natural Keys, Surrogate Keys, Entity Resolution and Identity Drift — connect sequence extraction to the later identity decisions that consume its spans.
• Indicator Validation, Base Rates and Warning Thresholds — treat model outputs as indicators requiring base-rate and threshold evidence.
• Schema Evolution, Backward Compatibility and Semantic Drift — govern changes in label ontologies and downstream meaning.

Questions readers often ask

How does a CRF differ from independent classification?

It scores a whole label structure and can model dependencies between output labels rather than classifying each position separately.

How does a CRF differ from an MRF?

A CRF directly models a conditional distribution of labels given observations; MRF is the broader undirected-model family and may model a joint distribution.

Does Viterbi output provide label probabilities?

No. It gives the best complete sequence under the model. Forward-backward provides marginals for individual positions or transitions.

Do hard BIO constraints solve annotation quality?

No. They prevent invalid paths but cannot settle inconsistent definitions or ambiguous span boundaries.

Does a neural encoder make CRF features unnecessary?

It changes how evidence is represented, but transition assumptions, labels, evaluation, drift and governance still require explicit checks.

Can a coherent sequence still be wrong?

Yes. Structured consistency is not truth; the data, label policy, domain shift and model uncertainty can all produce confident errors.

Sources and scope

Lafferty, McCallum and Pereira: Conditional Random Fields — the original 2001 paper introducing CRFs for segmenting and labelling sequence data.

Sutton and McCallum: An Introduction to Conditional Random Fields — comprehensive tutorial on structure, inference, training and implementation.

An Introduction to Conditional Random Fields, full text — technical reference for derivations and practical considerations.

Sha and Pereira: Shallow Parsing with Conditional Random Fields — primary application paper demonstrating sequence labelling and decoding.

McCallum: Efficiently Inducing Features of Conditional Random Fields — research on feature induction and complexity control for CRFs.

The library requests, labels and service decisions in the worked examples are fictional. This guide is educational and does not authorise automated high-stakes decisions, identity claims or service denial without lawful purpose, validation, human review and remedy.

“Gold standard” describes the educational benchmark explained in this guide. It is not an official accreditation, legal opinion, policy determination or professional certification.

Back to contents · Return to the Judgement root · Return to the Gold Standard of Education