Graph-Based Reranking

Graph-based reranking operates as a cascade architecture where an initial semantic retrieval pass is refined by structural graph signals:

• Stage 1 — Semantic Retrieval: A dense embedding model retrieves an initial candidate set of N documents based on topical similarity to the query
• Stage 2 — Graph Reranking: Each candidate's position is recalculated using authority scores derived from the citation graph, such as PageRank variants or betweenness centrality
• The final ranked list balances topical relevance with precedential weight, preventing highly cited but topically irrelevant cases from dominating
• This decoupling allows independent optimization of the semantic encoder and the graph authority model

The core mechanism of graph-based reranking is the injection of precedential influence metrics into the relevance calculation:

• PageRank Variants: Eigenvector centrality algorithms adapted for directed citation graphs, where citations function as votes of authority
• Weighted Edges: Citation edges are weighted by treatment type — a 'followed' citation carries more authority than a 'distinguished' or 'criticized' reference
• Jurisdictional Filtering: Authority scores are constrained to the relevant court hierarchy, preventing a persuasive authority from a foreign jurisdiction from outranking binding circuit precedent
• The final score is typically a linear combination or learned fusion of semantic similarity and graph-derived authority

Not all citations are equal. Graph-based reranking incorporates citation sentiment and treatment classification to modulate authority propagation:

• Positive Treatment: 'Followed,' 'applied,' or 'affirmed' citations strengthen the cited node's authority score
• Negative Treatment: 'Overruled,' 'criticized,' or 'questioned' citations reduce or nullify authority propagation along that edge
• Neutral Treatment: 'Cited,' 'discussed,' or 'explained' references contribute minimal weight
• This creates a signed, weighted graph where authority flows preferentially through positive citation paths, preventing bad law from achieving high reranking scores

Legal authority is not static — it evolves over time. Graph-based reranking systems incorporate temporal dynamics to prevent outdated precedents from dominating results:

• Recency Weighting: More recent citations contribute higher authority propagation weights, reflecting the legal system's preference for current interpretations
• Precedent Aging Curves: Cases that have not been cited in extended periods receive diminishing authority scores, modeling the concept of desuetude
• Overruling Cascades: When a seminal case is overruled, the negative treatment signal propagates downstream through its precedent chain, downgrading all dependent authorities
• Temporal snapshots enable point-in-time reranking, critical for analyzing legal questions as they stood at a specific date

The fusion of semantic and graph signals requires careful calibration. Common hybrid scoring approaches include:

• Linear Interpolation: FinalScore = α × SemanticSimilarity + (1-α) × NormalizedAuthority, where α is tuned on held-out relevance judgments
• Learning-to-Rank: A supervised model trained on annotated legal research queries learns the optimal weighting of semantic features, graph centrality metrics, and treatment signals
• Reciprocal Rank Fusion (RRF): Combines the rank positions from separate semantic and graph-based rankings without requiring score calibration, effective when score distributions differ significantly
• Cascade Filtering: Graph authority is applied as a hard filter rather than a soft reranker — only candidates exceeding an authority threshold survive the second stage

Citation networks naturally cluster into doctrinal communities — groups of cases addressing related legal questions. Graph-based reranking leverages this structure:

• Community Detection: Algorithms like Louvain or Leiden partitioning identify clusters of densely interconnected cases representing distinct legal topics
• Intra-Community Boost: Candidates within the same community as the top semantic results receive authority score amplification, reinforcing topical coherence
• Cross-Community Penalization: Highly authoritative cases from unrelated doctrinal clusters are deprioritized, solving the problem where a landmark constitutional case ranks highly for a contracts query
• This ensures reranking respects topical boundaries within the citation graph rather than applying a global authority score indiscriminately

A directed network structure where nodes represent legal cases or statutes and edges represent citation relationships. This forms the foundational data structure for computational precedent analysis. In graph-based reranking, the citation graph provides the topological substrate over which authority propagation algorithms operate to reorder initial semantic search results.

A graph algorithm that iteratively distributes precedential influence scores across a citation network, often using PageRank variants like Eigenfactor or weighted PageRank. In a two-stage retrieval pipeline, these scores are used to boost or demote documents from an initial BM25 or semantic search result set, ensuring that legally influential cases surface to the top regardless of keyword density.

A quantitative metric computed over a citation graph that estimates the precedential weight of a legal case. It is calculated based on:

• In-degree centrality: number of citing cases
• Citation quality: authority of citing sources
• Treatment sentiment: whether citations are positive or negative

In reranking, documents with higher authority scores are promoted to prioritize legally influential precedent.

The architectural pattern where an initial fast retriever (e.g., BM25 or dense embeddings) generates a candidate set of documents, and a slower, more precise reranker reorders them. Graph-based reranking is a specific reranking strategy that uses citation network topology as the scoring signal, combining semantic relevance with structural legal authority to produce a final ranked list.

A graph metric measuring how often a node lies on the shortest path between other nodes. In citation networks, high betweenness centrality identifies cases that serve as critical bridges connecting distinct doctrinal clusters. These bridge cases are often seminal decisions that synthesize multiple lines of precedent, making them high-value targets for reranking in multi-document legal reasoning tasks.

A deep learning architecture designed to operate directly on graph-structured data. In legal AI, GNNs learn node embeddings that capture both a case's intrinsic features and its citation neighborhood structure. These learned embeddings can serve as sophisticated authority signals for reranking, going beyond simple centrality metrics to capture complex, non-linear patterns of legal influence.