Active Learning

The most common query strategy where the model selects examples it is least confident about for human annotation.

• Least Confidence: Selects instances where the model's highest predicted probability is lowest.
• Margin Sampling: Chooses examples where the difference between the top two predicted class probabilities is smallest, indicating decision boundary ambiguity.
• Entropy Sampling: Picks instances with the highest entropy across all possible label predictions, maximizing information gain.

For token-level NER tasks, uncertainty is often aggregated across all tokens in a sequence using token entropy or sequence probability.

A strategy that maintains an ensemble of models trained on the same labeled data but with different initializations or architectures. The committee votes on predictions for unlabeled examples.

• Instances with the highest disagreement among committee members are selected for annotation.
• Disagreement is measured using vote entropy or Kullback-Leibler divergence between model output distributions.
• QBC reduces the risk of selecting outliers that a single model might find confusing but are not representative of the true data distribution.

Prevents the selection of redundant or highly similar examples by incorporating a diversity constraint into the query process.

• Clustering-based: Embed unlabeled examples using the model's encoder, cluster them, and sample uncertain instances from different clusters.
• Determinantal Point Processes (DPP): A probabilistic model that selects a subset of examples that are both individually informative and collectively diverse.
• Core-set selection: Formulates the problem as finding a subset of unlabeled data that best represents the entire unlabeled pool in the embedding space.

Diversity sampling is critical for avoiding batch-mode redundancy when querying multiple examples simultaneously.

The initial phase of active learning where no labeled data exists to train a seed model, making uncertainty-based querying impossible.

• Random sampling is the default strategy for the first batch, but it is inefficient.
• Clustering-based initialization selects diverse examples from the unlabeled pool for the first annotation round.
• Transfer learning from a pre-trained model on a related domain can provide initial uncertainty estimates.
• Domain expert heuristics or gazetteers can identify high-value seed examples likely to contain entities.

Effective cold start strategies can reduce annotation costs by up to 40% compared to random initialization.

Determining when the active learning loop has achieved sufficient performance to terminate annotation, balancing cost against marginal accuracy gains.

• Performance plateau detection: Stop when F1 score improvement over the last k iterations falls below a threshold (e.g., < 0.5% absolute F1).
• Annotation budget exhaustion: A fixed-cost approach where labeling stops after a predetermined number of examples or budget is spent.
• Confidence-based stopping: Terminate when the model's average confidence on the remaining unlabeled pool exceeds a threshold (e.g., 95%).
• Inter-annotator agreement: Stop when the model's predictions agree with human annotators at a rate comparable to human-human agreement.

A practical variant that selects multiple examples simultaneously for annotation in each iteration, rather than querying one at a time.

• Addresses the latency of sequential human annotation workflows where annotators work in parallel.
• Requires balancing informativeness and diversity within each batch to avoid selecting redundant examples.
• BatchBALD extends the Bayesian Active Learning by Disagreement framework to select batches that maximize mutual information between model parameters and batch labels.
• Typical batch sizes range from 10 to 100 examples per iteration, depending on annotator capacity and model retraining cost.

The most common query strategy in active learning. The model identifies examples where it is least confident about the prediction.

• Least Confidence: Queries the instance with the lowest predicted probability for its most likely label.
• Margin Sampling: Queries the instance with the smallest difference between the top two predicted class probabilities, indicating a decision boundary conflict.
• Maximum Entropy: Queries the instance where the predicted probability distribution over all possible entity tags has the highest entropy, reflecting maximal disorder.

A disagreement-based strategy that maintains an ensemble of distinct NER models trained on the same labeled data. The system queries the instance where the committee of models disagrees most on the output label sequence.

• Vote Entropy: Measures the disagreement using the entropy of the label votes across the committee.
• KL Divergence: Queries the instance that maximizes the average Kullback-Leibler divergence between any single committee member and the consensus.

A hybrid strategy that prevents the selection of outliers. It balances informativeness (uncertainty) with representativeness (density).

• The model scores unlabeled instances based on their similarity to other instances in the feature space.
• Prevents the oracle from wasting time labeling noisy or highly idiosyncratic examples that won't generalize well to the rest of the corpus.

The initial seed set selection problem before any labels exist. Since there is no trained model to measure uncertainty, the system must use unsupervised heuristics to pick the first batch for annotation.

• K-Means Clustering: Selects instances closest to cluster centroids in the embedding space.
• Core-Set Selection: A greedy algorithm that picks a subset of data points such that the distance from any unlabeled point to the nearest labeled point is minimized.

The software layer that connects the active learning algorithm to the human oracle. Effective HITL design is critical for reducing annotation latency and cognitive load.

• Pre-Annotation: The model provides its best guess as a starting point, which the human corrects rather than labeling from scratch.
• Bulk Annotation: Grouping similar or related instances to leverage context and speed up throughput.
• Active Cleaning: Flagging and re-labeling suspected errors in the existing training set, not just labeling new data.

The logic that determines when the active learning loop has achieved sufficient performance and can terminate to avoid unnecessary annotation costs.

• Performance Plateau: Stop when the F1 score on a held-out validation set ceases to improve by a significant delta over consecutive iterations.
• Confidence Convergence: Stop when the average model confidence on the remaining unlabeled pool exceeds a high threshold.
• Budget Exhaustion: A hard stop triggered by the depletion of a pre-allocated annotation budget.