Node Classification

The foundational mechanism of Graph Neural Networks (GNNs) where nodes iteratively aggregate feature information from their local neighborhood. Each node updates its hidden state by combining its own features with those of its neighbors, enabling the model to capture both attributive and structural patterns. In legal graphs, this allows a 'case' node to refine its representation based on connected 'statute', 'court', and 'citation' nodes before classification.

A spectral-based GNN variant that performs convolution operations directly on graph structures. GCNs use a symmetric normalization of the adjacency matrix to weight neighbor contributions, preventing high-degree nodes from dominating the learning signal. For legal knowledge graphs, GCNs effectively classify document types (e.g., distinguishing a 'dissenting opinion' from a 'majority opinion') by leveraging the homophily assumption—that connected legal documents tend to share similar characteristics.

An attention-based architecture that assigns learnable importance weights to different neighbors during aggregation, rather than treating all connections equally. This is critical for legal citation networks where some precedents are highly authoritative while others are merely persuasive. GATs allow the model to dynamically prioritize binding citations from superior courts over tangential references, producing more nuanced node embeddings for tasks like classifying a case's legal domain.

A classical semi-supervised technique that propagates labels through the graph based on edge connectivity without learning node embeddings. Labels spread from seeded nodes to unlabeled neighbors iteratively until convergence. In legal knowledge graphs, this provides a strong baseline for tasks like classifying statutory sections into legal topics, assuming that closely connected sections in a code hierarchy share subject matter. Computationally lightweight compared to GNNs.

An extension of GCNs designed for multi-relational graphs where edges have distinct types. R-GCNs learn separate transformation matrices for each relationship type (e.g., 'cites', 'overturns', 'applies', 'distinguishes'). This is essential for legal knowledge graphs where the semantic meaning of a connection dramatically affects node classification—a case 'overturned by' a higher court carries fundamentally different information than a case merely 'cited by' a peer court.

Techniques for classifying nodes when only a handful of labeled examples exist per category. Methods include meta-learning (training models to rapidly adapt to new classes) and prototypical networks that compute class centroids in embedding space. In legal contexts, this addresses the scarcity of annotated data for niche legal doctrines—enabling classification of rare procedural postures or emerging areas of law where exhaustive manual labeling is impractical.

A deep learning architecture that operates directly on graph structures. GNNs use message passing between nodes to capture dependencies and generate representations. In legal contexts, GNNs can learn to classify case nodes by aggregating features from neighboring citations and statutory references.

• Message Passing: Nodes iteratively exchange feature information with neighbors
• Inductive Learning: Generalizes to unseen nodes without retraining
• Legal Application: Predicting case outcomes based on citation network topology

A technique that maps nodes, edges, and their features into a low-dimensional continuous vector space while preserving structural and relational properties. These embeddings serve as input features for downstream node classification tasks.

• Transductive methods like Node2Vec use random walks to capture local topology
• Inductive methods like GraphSAGE learn aggregation functions
• Legal entities with similar citation patterns occupy proximate positions in embedding space

A graph analytics task that estimates the likelihood of a missing or future connection between two nodes. In legal knowledge graphs, link prediction identifies unrecorded citations, relevant precedents, or hidden relationships between cases and statutes.

• Uses learned node embeddings to score potential edges
• Heuristic methods: Common Neighbors, Adamic-Adar
• Learned methods: GNN-based encoders with edge decoders
• Critical for knowledge base completion in legal domains

The Resource Description Framework is a W3C standard that structures information as subject-predicate-object triples forming directed, labeled graphs. Legal knowledge graphs often use RDF to represent entities like cases, statutes, and parties with formal semantics.

• Triple format: (CaseA, cites, CaseB)
• Enables SPARQL querying across distributed legal datasets
• Foundation for OWL ontologies that define legal concept hierarchies

The SPARQL Protocol and RDF Query Language is a W3C-standardized query language for retrieving and manipulating data stored in RDF format. It enables precise traversal of legal knowledge graphs to extract subgraphs for node classification training.

• Pattern matching against triple structures
• Supports federated queries across multiple legal triplestores
• Used to construct training datasets by selecting labeled nodes and their k-hop neighborhoods

The NLP task of connecting textual entity mentions to their unique, unambiguous identifiers in a knowledge base. In legal pipelines, NEL maps references like 'Brown v. Board of Education' to a specific node URI before classification occurs.

• Resolves entity ambiguity across different naming conventions
• Essential preprocessing step before node feature extraction
• Links unstructured legal text to structured graph nodes for semi-supervised learning