Citation Recommendation

Modern systems combine dense semantic retrieval with graph-based reranking to overcome the limitations of each approach alone. The process operates in two stages:

• First-pass retrieval: A legal embedding model encodes the query brief into a vector, retrieving the top-k semantically similar documents from a vector database
• Second-pass reranking: Retrieved candidates are reordered using authority scores and graph centrality metrics from the citation network, ensuring legally influential cases surface above merely topically similar ones

This hybrid approach prevents the system from recommending factually similar but legally irrelevant or overruled authorities.

Citation recommenders weight candidates using precedential influence scores derived from the citation graph rather than relying solely on textual similarity. Key signals include:

• Binding vs. persuasive authority: The system applies jurisdictional filtering to prioritize mandatory precedent from higher courts within the same jurisdiction
• Treatment type classification: Cases flagged with negative treatment or overruling detection are demoted or excluded
• Authority propagation: Graph algorithms like PageRank variants distribute influence scores across the network, identifying seminal cases with sustained citation velocity

This ensures recommendations reflect legal weight, not just topical relevance.

Advanced recommenders classify the rhetorical purpose of a citation to improve suggestion quality. Citation intent classification distinguishes between:

• Legal support: The cited authority provides direct precedential backing for a proposition
• Factual analogy: The case shares materially similar facts
• Background context: The citation establishes general legal principles
• Critical disagreement: The citing court distinguishes or criticizes the authority

By modeling intent, the system can match the drafter's rhetorical need—suggesting supportive cases when building an argument and distinguishing cases when anticipating counterarguments.

Citation recommenders incorporate temporal citation analysis to model how legal authority evolves over time. Critical temporal signals include:

• Precedent aging: Older cases may lose relevance if subsequent decisions have refined or narrowed their holdings
• Citation cascades: A seminal decision triggering a chain reaction of citations indicates growing doctrinal importance
• Recency weighting: Recent decisions interpreting a statute are often more relevant than older ones, particularly in rapidly evolving regulatory areas

Temporal modeling prevents the system from recommending historically important but practically superseded authorities.

Cutting-edge systems employ Graph Neural Networks (GNNs) to learn node embeddings that capture both a case's intrinsic textual features and its structural position within the citation graph. GNNs enable:

• Neighborhood-aware representations: A case's embedding reflects the collective influence of citing and cited authorities
• Heterogeneous graph modeling: Systems simultaneously model cases, statutes, courts, and judges as distinct node types with typed edges representing different relationships
• Link prediction: GNNs can forecast which precedents a court is likely to cite, enabling proactive recommendation before a citation pattern becomes obvious

This approach moves beyond static authority scores to learned, context-sensitive representations.

Community detection algorithms partition the citation graph into clusters of densely interconnected cases, often corresponding to distinct legal doctrines or circuits. Recommendation systems leverage these clusters to:

• Expand doctrinal coverage: Once a drafter cites one case from a doctrinal cluster, the system surfaces other influential cases within the same community
• Identify seminal cases: Nodes with high betweenness centrality that bridge multiple communities are flagged as foundational authorities
• Detect circuit splits: Distinct community structures across different federal circuits can highlight conflicting interpretations requiring resolution

Community awareness ensures recommendations are doctrinally coherent rather than topically scattered.

A graph algorithm that iteratively distributes precedential influence scores across a citation network. Similar to PageRank, it identifies the most legally significant nodes by analyzing both direct citations and the authority of citing sources.

• Recursive scoring: A case is authoritative if cited by authoritative cases
• Handles binding vs. persuasive weight distinctions
• Dampens influence across jurisdictional boundaries
• Produces rankable authority scores for retrieval reranking

A two-stage retrieval technique where an initial semantic search result set is reordered using graph centrality or authority scores from a citation network.

• Stage 1: Dense retrieval using legal embeddings for semantic similarity
• Stage 2: Rerank candidates by betweenness centrality or authority score
• Prioritizes legally influential documents over merely topically similar ones
• Combines semantic relevance with structural precedent importance

A fine-grained NLP task that determines the rhetorical purpose of a citation. Understanding why a case was cited—not just that it was cited—enables more nuanced recommendation.

• Supportive: Citing as binding or persuasive authority
• Analogical: Drawing factual parallels
• Background: Providing legal context
• Critical: Disagreeing or distinguishing
• Enables filtering recommendations by intended rhetorical function

A machine learning task that predicts the likelihood of a missing or future edge between two nodes in a citation graph. Applied to legal networks to forecast which precedents a court is likely to cite.

• Uses Graph Neural Networks (GNNs) to learn node embeddings
• Incorporates temporal features for evolving citation patterns
• Enables proactive suggestion of relevant but uncited authorities
• Critical for identifying gaps in legal argumentation

An NLP task that automatically categorizes how a citing case legally treats a cited authority. Essential for maintaining accurate authority graphs and filtering recommendations.

• Overruled: Explicitly invalidated
• Distinguished: Found materially different
• Followed: Applied as controlling precedent
• Criticized: Questioned but not overruled
• Prevents recommending weakened or invalidated authorities