Treatment Type Classification

A controlled vocabulary of categorical signals assigned to each citation instance. Standard labels include:

• Overruled: The cited authority's holding is expressly invalidated.
• Distinguished: The court finds material factual or legal differences, declining to apply the precedent.
• Followed: The court adopts and applies the cited reasoning.
• Criticized: The court questions the soundness of the cited decision without overruling it.
• Harmonized: The court reconciles an apparent conflict between authorities.
• Limited: The scope of the cited authority is explicitly narrowed.

Treatment classification operates at multiple levels of specificity:

• Binary Polarity: Simple positive vs. negative treatment detection.
• Multi-Class Categorization: Assignment to one of 6-12 discrete treatment types.
• Fine-Grained Subtype Analysis: Distinguishing between 'expressly overruled' and 'impliedly overruled'.
• Multi-Label Classification: A single citation may simultaneously 'follow' a case for one proposition while 'distinguishing' it on another point.
• Span-Level Annotation: Pinpointing the exact textual segment within a citing opinion that conveys the treatment signal.

Key linguistic and structural features used by models to predict treatment:

• Citational Context Window: The surrounding sentences and paragraphs where the citation string appears.
• Discourse Markers: Explicit signals like 'we decline to follow,' 'the dissent criticizes,' or 'as the court correctly held.'
• Court Hierarchy Delta: The relative jurisdictional level between citing and cited courts.
• Temporal Distance: The time elapsed between the cited and citing decisions.
• Subsequent Treatment History: The existing treatment signals from other citing cases in the graph.

Modern approaches leverage domain-specific language models:

• Fine-Tuned Legal Encoders: Models like Legal-BERT or CaseLaw-BERT fine-tuned on annotated treatment datasets.
• Graph Neural Networks (GNNs): Architectures that incorporate both the textual context and the structural position of the citation within the broader authority graph.
• Multi-Task Learning: Jointly training on treatment classification, citation intent, and citation sentiment to leverage shared representations.
• Few-Shot Prompting: Using large language models with carefully engineered prompts containing exemplar citation contexts and their treatment labels.

Accurate treatment classification is foundational for:

• Shepardizing Automation: Replacing manual citator review with real-time, algorithmic validity checks.
• Authority Score Computation: Weighting citation graph edges by treatment polarity to compute nuanced precedential influence scores.
• Overruling Event Detection: Triggering alerts when a seminal case receives a negative treatment signal from a higher court.
• Precedent Chain Integrity: Ensuring that automated reasoning chains do not rely on authorities that have been subsequently criticized or overruled.

Rigorous evaluation requires legal expertise:

• Cohen's Kappa: Measures inter-annotator agreement between human legal experts, establishing the task's ceiling.
• Macro F1-Score: Evaluates performance across all treatment classes, critical for handling imbalanced label distributions where 'followed' dominates.
• Confusion Matrix Analysis: Identifies systematic model errors, such as confusing 'criticized' with 'distinguished'.
• Blind Citator Comparison: Benchmarking model output against the treatment signals assigned by commercial citator services like Shepard's or KeyCite.

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.

• Directed edges capture the flow of authority from citing case to cited case
• Node attributes include court level, jurisdiction, and date
• Edge attributes can encode treatment type and sentiment
• Serves as the input graph for authority propagation algorithms like PageRank variants

The process of using a citator service to trace a legal authority's subsequent treatment history. This determines whether a case remains good law and has not been overruled, criticized, or otherwise negatively treated.

• Originates from Shepard's Citations, first published in 1873
• Modern computational Shepardizing uses treatment type classification to automate signal assignment
• Critical for negative treatment detection and overruling detection
• Outputs inform precedential weight calculations in authority graphs

A graph algorithm that iteratively distributes precedential influence scores across a citation network. Often implemented using PageRank variants, it identifies the most legally significant nodes.

• Weighted edges incorporate citation sentiment and treatment signals
• Jurisdictional filtering constrains propagation to legally relevant courts
• Outputs feed into authority score and precedent influence score metrics
• Enables seminal case detection by surfacing high-centrality origin nodes

A fine-grained NLP task that determines the rhetorical purpose of a citation instance. Goes beyond binary sentiment to classify whether a case is cited for legal support, factual analogy, background context, or critical disagreement.

• Distinct from treatment type classification, which focuses on legal effect
• Uses transformer-based models fine-tuned on annotated legal corpora
• Enables richer edge typing in heterogeneous graphs
• Improves accuracy of citation recommendation and graph-based reranking

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.

• Message passing aggregates information from neighboring nodes
• Used for link prediction to forecast future citations
• Enables community detection to identify doctrinal clusters
• Applied to case outcome prediction by encoding precedent relationships

The study of citation patterns over time to model how legal authority evolves, ages, or gains influence. Incorporates timestamps into graph models to detect trends like precedent aging and citation cascades.

• Citation cascades trace the propagation of a seminal decision through time
• Precedent aging models the decay of authority for older decisions
• Enables detection of emerging doctrinal shifts before they become explicit
• Critical for accurate stare decisis modeling in dynamic legal landscapes