Graph Neural Network Linking

Unlike local models that resolve each mention independently, GNN linking performs collective inference. The model evaluates the compatibility of all candidate entities in a document simultaneously. If one mention is linked to 'Mercury' the planet, the GNN propagates a signal that makes 'Atmosphere' more likely to link to a planetary concept than a unit of pressure. This global coherence constraint dramatically reduces errors in dense technical text.

The core mechanism involves message passing along typed edges in the knowledge graph. Each entity node aggregates feature vectors from its neighbors, weighted by the relationship type. For example, a 'Drug' node receives distinct messages from 'has_ingredient' edges versus 'treats_disease' edges. This allows the model to learn that a mention of 'insulin' in the context of 'pancreas' should favor the hormone concept over the pharmaceutical product.

GNNs enable multi-hop reasoning across the knowledge graph. The model can traverse multiple edges to find supporting evidence for a disambiguation decision. A mention of 'JAK2' in a text discussing myeloproliferative disorders can be linked correctly by traversing the path: JAK2 → associated_with → Myeloproliferative Disorder → subClass_of → Bone Marrow Disease. This structural context is inaccessible to pure text-based encoders.

The architecture jointly embeds the textual mention context and the knowledge graph structure into a unified vector space. A transformer encoder processes the surrounding clinical text, while a GNN encoder processes the subgraph of candidate entities. A final scoring layer computes compatibility scores between each mention's contextualized embedding and every candidate's graph-informed embedding, selecting the most coherent global configuration.

Biomedical knowledge graphs contain diverse node types: Diseases, Drugs, Genes, Procedures, and Anatomical Sites. GNN linkers explicitly model these types as distinct node categories with separate projection matrices. This prevents the model from confusing structural patterns across domains—a gene's interaction network has different topological signatures than a disease hierarchy, and the model learns to respect these ontological boundaries.

Modern GNN linkers replace simple averaging with graph attention mechanisms. During message passing, the model learns to assign higher weight to relevant neighbors and down-weight noisy connections. If a mention of 'hypertension' has candidates connected to both 'heart failure' and 'skin rash' in the context, the attention heads will amplify the cardiovascular pathway and suppress the dermatological one, yielding contextually sharp disambiguation.

The core mechanism by which GNNs operate, where each node in the knowledge graph iteratively aggregates feature vectors from its local neighborhood. In entity linking, a mention node receives information from connected concept nodes, synonym nodes, and co-occurring mention nodes. This neighborhood aggregation allows the model to refine a mention's representation based on the relational structure of the graph, effectively using the homophily principle—the tendency for linked entities to be semantically related—to disambiguate terms. Frameworks like GraphSAGE and Graph Attention Networks (GATs) implement this with learnable aggregation functions.

Unlike local models that link each mention independently, GNN-based approaches perform collective inference by jointly modeling all mentions in a document. The graph connects co-occurring mentions with edges representing semantic coherence or co-occurrence statistics. This allows the model to enforce global consistency: if 'diabetes' is linked to a concept in a metabolic disorder subtree, the model is biased to link 'neuropathy' to diabetic neuropathy rather than an unrelated form. This global normalization significantly outperforms pipelined, mention-by-mention approaches on ambiguous clinical narratives.

A specialized GNN architecture designed to handle heterogeneous knowledge graphs with multiple edge types. In clinical entity linking, edges represent distinct relationships: has_synonym, has_subtype, may_be_caused_by, has_finding_site. R-GCNs apply separate weight matrices for each relation type during message passing, allowing the model to learn that information flowing through a hierarchical 'isa' edge should be treated differently than information from a causal 'may_be_caused_by' edge. This relational awareness is critical for modeling the rich semantics of ontologies like SNOMED CT.

GATs introduce a self-attention mechanism to the message passing process, allowing the model to learn the importance of each neighboring node dynamically. Instead of aggregating all neighbors equally, a GAT computes attention coefficients that weight the contribution of each edge. For entity linking, this means the model can learn to attend more strongly to a synonym node that provides strong lexical evidence than to a distant sibling concept node. This selective focus improves disambiguation precision, especially when the local graph neighborhood is large and noisy.

The quality of initial node features critically impacts GNN linking performance. Concept nodes are typically initialized with dense vectors from pre-trained biomedical language models like SapBERT or PubMedBERT, which encode the textual descriptions of each concept. Mention nodes are initialized with contextual embeddings from the same model, encoding the mention's surface form and surrounding clinical context. This semantic initialization provides a strong prior, allowing the GNN to focus on refining representations through relational structure rather than learning semantics from scratch.

A training paradigm that directly optimizes the GNN to produce similar embeddings for a mention and its correct concept while pushing apart embeddings for incorrect candidates. The loss function operates on the final node representations after message passing. Hard negative mining is essential: the model is trained to distinguish the true concept from highly confusable alternatives that share the same semantic type or lexical features. This fine-grained discriminative training significantly improves the model's ability to resolve subtle disambiguation cases.