Nested NER

The defining characteristic of nested NER is the ability to model parent-child relationships between entity spans. Unlike flat NER, which assigns a single label to a token, nested systems recognize that a text segment can simultaneously belong to multiple semantic categories at different levels of granularity.

• Example: In 'metastatic carcinoma of the lung', the entire phrase is a Diagnosis, while 'lung' is a Body_Site and 'carcinoma' is a Histology.
• Clinical Relevance: This mirrors how clinicians document—moving from broad impressions to specific anatomical and morphological details within a single phrase.

Nested NER systems must distinguish between two structural patterns:

• Strict Nesting: One entity is fully contained within another. The inner entity's span is a complete substring of the outer entity's span.
• Overlapping Entities: Two entities share some tokens but neither fully contains the other. This is common in coordinate structures like 'liver and kidney failure', where 'liver' and 'kidney' are separate Body_Site entities overlapping with the Diagnosis.

Handling both requires architectures that move beyond simple BIO tagging, which assumes non-overlapping, sequential labels.

Standard linear-chain CRFs cannot model nested structures because they assume a single tag per token. Nested NER requires specialized decoding strategies:

• Layered Approach: Deploy multiple stacked NER models, each predicting entities at a different level of granularity (e.g., one layer for Medication, another for Dosage).
• Span-Based Classification: Enumerate all possible text spans up to a maximum length and classify each span independently as an entity type or 'none'. This naturally handles nesting and overlapping.
• Pointer Networks: Use a generative approach where the model predicts the start and end indices of entities, allowing multiple entities to share tokens.

An alternative to span classification is treating nested NER as a shift-reduce parsing problem, adapted from constituency parsing in NLP.

• Mechanism: The model reads tokens sequentially and uses a stack to build a forest of entity trees. Actions like SHIFT, REDUCE-ENTITY, and UNARY construct nested structures incrementally.
• Advantage: This method explicitly models the hierarchical dependencies between entities, ensuring that a Dosage cannot be created without a parent Medication.
• Clinical Application: Enforces ontological constraints, such as a Severity modifier being attached to a Finding rather than existing independently.

Standard token-level F1 score is insufficient for evaluating nested NER because it penalizes systems that correctly identify an entity but assign it to the wrong hierarchical level. Specialized metrics are required:

• Entity-Level Precision/Recall: An entity is considered a true positive only if both its span boundaries and entity type match exactly.
• Strict vs. Lenient Matching: Strict matching requires exact boundary match; lenient matching may give partial credit for overlapping spans.
• Tree Edit Distance: Measures the structural similarity between the predicted entity tree and the gold-standard tree, capturing hierarchical errors that flat metrics miss.

Manually annotating nested clinical entities is exponentially more expensive and error-prone than flat annotation. A common mitigation strategy is synthetic data generation:

• Template-Based Composition: Programmatically combine known entities into nested structures using clinical templates. For example, inserting a known Drug and Dosage into a Medication phrase template.
• Context-Free Grammar (CFG): Define a grammar that generates valid nested clinical expressions, ensuring coverage of rare nesting patterns.
• Data Augmentation: Apply back-translation or synonym replacement to existing nested annotations to increase dataset diversity without manual effort.

A modern NER paradigm that directly classifies arbitrary text spans as entity types, bypassing token-level BIO tagging. This approach is inherently superior for nested NER because it can enumerate all possible spans in a sentence and classify each one independently, naturally capturing overlapping and hierarchically contained entities like a 'dosage' inside a 'medication' mention without conflicting tag sequences.

The standard token-level annotation scheme using B (Beginning), I (Inside), and O (Outside) tags. While foundational for flat NER, standard BIO tagging fundamentally breaks with nested entities because a single token cannot simultaneously be the beginning of one entity and the inside of another. Extended schemes like BILOU or multi-layered BIO are often required as workarounds for hierarchical clinical concepts.

A discriminative probabilistic model for segmenting and labeling sequence data. CRFs excel at capturing dependencies between adjacent tags, ensuring valid BIO transitions. However, standard linear-chain CRFs assume a single label per token, making them incompatible with nested structures. They are often replaced by hypergraph-based CRFs or discarded entirely in favor of span-based models for nested clinical NER tasks.

The downstream task of mapping a recognized clinical entity mention to its unique Concept Unique Identifier (CUI) in the UMLS Metathesaurus. In nested NER, normalization becomes more complex because a parent entity (e.g., 'metastatic carcinoma') and its child (e.g., 'carcinoma') may map to different CUIs with distinct semantic types, requiring the system to maintain hierarchical grounding fidelity.

A modular, sequential architecture orchestrating components for tokenization, NER, negation detection, and concept normalization. Supporting nested NER requires a pipeline design that preserves span provenance across stages—ensuring that a negation signal applied to a parent entity correctly propagates to all its children, and that extracted hierarchical relationships survive through to final structured output.

An approach combining high-precision rule-based systems with high-recall statistical models. For nested entities, hybrid systems can use deterministic rules to define parent-child relationships (e.g., a regex for 'dosage' always being a child of 'medication') while letting a neural span categorizer handle ambiguous boundaries. This balances the structural rigidity needed for hierarchy with the flexibility of machine learning.