Few-Shot NER

Uses prototypical networks to compute a similarity function between a query token and a small support set of labeled entity examples. The model learns a metric space where tokens belonging to the same entity type cluster closely together, allowing classification via nearest-neighbor lookup rather than a fixed softmax layer. This enables immediate adaptation to new entity types without retraining.

Reformulates NER as a cloze-style language modeling task. A template like "[Text] The organization is [MASK]" is fed to a pre-trained language model, and the output at the masked position is decoded into the entity span. This leverages the vast parametric knowledge of models like BART or T5, allowing them to extract entities with as few as 5-10 examples by simply conditioning on the prompt template.

A model is first fully trained on a high-resource source domain with abundant labeled data, then fine-tuned on the target domain with only K examples per entity type. The source domain teaches general entity boundary detection and syntactic patterns, while the few-shot fine-tuning adapts the type classifier to the target schema. This is distinct from zero-shot transfer, as it still requires a minimal target-domain signal.

Employs contrastive objectives to pull representations of entity spans of the same type together while pushing apart spans of different types or non-entity text. In a few-shot setting, a support set of labeled spans defines the positive anchors. The model scores candidate spans from a query sentence against these anchors, selecting the highest-similarity match. This approach naturally handles nested and overlapping entities.

When only 5 or 10 examples are available, synthetic data generation becomes critical. Techniques include back-translation, synonym replacement, and contextual word substitution using masked language models to create semantically similar but lexically diverse training instances. This artificially expands the support set, preventing the model from overfitting to superficial lexical cues in the few provided examples.

Performance is measured by F1 score averaged over multiple few-shot episodes, not a single train/test split. Each episode samples a random K-shot support set from the target domain, and the model is evaluated on the remaining queries. Standard benchmarks include Few-NERD and CrossNER, which test both intra-domain and cross-domain few-shot generalization. Variance across episodes is reported to measure stability.

The process of adjusting a NER model trained on a source domain (e.g., news articles) to maintain high performance on a target domain (e.g., medical records) with distinct entity types and linguistic styles.

• Feature-level adaptation: Aligning input distributions across domains.
• Instance weighting: Emphasizing target-like examples in the source data.
• Parameter-efficient methods: Adapters and prefix tuning for rapid domain shift.

Few-shot NER is a specialized form of domain adaptation where the target domain provides only a handful of labeled instances.

A method for automatically generating noisy labeled training data by aligning a text corpus with an existing knowledge base or entity dictionary. This creates weak labels without manual annotation.

• Dictionary matching: Aligning text spans to Wikidata or DBpedia entries.
• Heuristic labeling functions: Pattern-based rules for rapid data generation.
• Noise mitigation: Combining distant labels with few-shot clean examples.

Distant supervision often provides the pre-training signal that few-shot NER models fine-tune from, creating a bridge between zero-resource and fully-supervised regimes.

A programmatic approach to generating training labels using multiple noisy heuristic functions—such as pattern matching, gazetteers, and regular expressions—managed by a generative model.

• Snorkel framework: Combines labeling functions into probabilistic training data.
• Label aggregation: Resolves conflicts between overlapping heuristics.
• Few-shot synergy: A small set of clean labels calibrates noisy supervision sources.

Weak supervision formalizes the creation of silver-standard data, enabling few-shot NER systems to bootstrap from minimal human effort while maintaining quality.

An iterative training strategy where a NER model intelligently queries a human annotator to label only the most informative and uncertain examples, minimizing annotation cost.

• Uncertainty sampling: Selecting instances where the model is least confident.
• Diversity sampling: Ensuring coverage across entity types and contexts.
• Cold-start problem: Initial seed selection for few-shot scenarios.

Active learning directly complements few-shot NER by optimizing which examples receive the scarce annotation budget, maximizing performance per labeled instance.

The task of assigning very specific semantic types from a large, hierarchically organized ontology—moving beyond coarse categories like PERSON or LOCATION to types like "American jazz pianist" or "biotechnology startup."

• Type hierarchies: Tree-structured labels with inheritance relationships.
• Zero-shot types: Generalizing to unseen fine-grained categories.
• Few-shot challenge: Sparse examples per type demand strong transfer learning.

FET represents the extreme end of the few-shot NER spectrum, where hundreds of types may each have only a handful of training instances.