Nested NER

Unlike flat NER, nested NER explicitly handles overlapping entity mentions where a shorter span is fully contained within a longer one. For example, in the phrase "[[University of [California]]] Berkeley", "California" is a Location nested inside the Organization "University of California Berkeley". This requires models to recognize multiple valid label paths for the same tokens.

Standard linear-chain CRF decoders cannot model nested structures. Nested NER architectures employ specialized decoding strategies:

• Layered-CRF: Stacks multiple flat NER layers, with each layer recognizing entities at a different nesting depth.
• Pyramid Models: Process text in a bottom-up fashion, connecting shorter spans into longer ones.
• Span-Based Classification: Enumerates all possible spans up to a maximum length and classifies each independently, naturally resolving overlaps.

A dominant paradigm for nested NER is to generate all possible text spans (up to a length constraint n) and classify them. Because the number of spans is O(n²), efficient filtering is critical. Techniques include:

• Span Pruning: Using a fast, low-cost scorer (e.g., a simple bilinear layer) to filter out the vast majority of background spans before feeding candidates to a heavier classifier.
• Biaffine Attention: Directly scoring all start-end token pairs in a single matrix operation to identify high-probability entity boundaries.

The standard benchmarks for nested NER are GENIA (biomedical domain) and ACE-2005 (news and broadcast). These datasets contain rich, deeply nested annotations. For instance, a single ACE-2005 sentence can have a Person entity nested inside an Organization entity, which is itself part of a Geopolitical Entity. Performance is measured using mention-level F1, requiring exact boundary and type matches for every nested layer.

A modern approach reformulates nested NER as a Machine Reading Comprehension task. For each entity type, a question is generated (e.g., "Which locations are mentioned?"). The model extracts answer spans directly from the text. Because multiple answers can be extracted independently for each question, this method naturally handles nested structures without complex decoding constraints, leveraging the power of pre-trained question-answering models.

Training nested NER models requires loss functions that account for the boundary sensitivity of overlapping spans. Standard cross-entropy on tokens is insufficient. Effective strategies include:

• Span-Level Cross-Entropy: Applied directly to the enumerated span representations.
• Dice Loss: Optimizes the overlap coefficient between predicted and ground-truth spans, directly targeting the mention-level F1 metric.
• Bipartite Matching Loss: Uses the Hungarian algorithm to find an optimal one-to-one assignment between predicted and gold entity spans before calculating the loss.

A NER paradigm that directly enumerates and classifies arbitrary text spans as entities, bypassing the need for token-level BIO tagging schemes. This approach is naturally suited for nested NER because it can independently score overlapping spans without the structural constraints of a linear-chain CRF. Architectures like Global Pointer and Biaffine Classifiers are common implementations.

A token-level annotation scheme where tokens are tagged as Beginning, Inside, or Outside of a named entity span. While foundational for flat NER, standard BIO tagging cannot represent nested entities because each token can only receive a single label. Extensions like BIOHD (with Head and Dependent tags) or multi-layered BIO schemes are required to encode hierarchical entity structures.

The task of assigning very specific semantic types from a large, hierarchically organized type ontology. FET and nested NER are deeply complementary: a nested NER system might identify 'San Francisco' as both a Location and a Settlement, while FET provides the deep type hierarchy (City → Municipality → PopulatedPlace) that defines the nesting structure.

The primary evaluation metric for NER that computes the harmonic mean of precision and recall based on exact matches of entity span boundaries and their type classifications. For nested NER, evaluation is more complex—metrics must account for hierarchical matches where a span is correctly identified but its parent-child relationship is wrong. Variants like loose F1 (partial boundary overlap) and strict F1 (exact span and type match) are commonly reported.

A segment-level sequence model that directly models the probability of entire entity spans rather than individual tokens. Unlike linear-chain CRFs, Semi-Markov CRFs can naturally handle segment-level features and are more amenable to nested structures by allowing overlapping segments with different labels. However, exact inference becomes computationally expensive for deep nesting hierarchies.

A NER architecture that uses a multiplicative attention mechanism to directly identify entity spans by scoring all possible start-end token pairs without a CRF layer. For nested NER, Global Pointer is particularly effective because it can predict multiple entity types independently—each type head scores all spans, allowing the same token to participate in multiple overlapping entities of different types.