Citation Sentiment

Citation sentiment is not a binary positive/negative flag but a graded spectrum of judicial attitude. Computational models typically classify citations into five to seven discrete categories.

• Strongly Positive: Explicit endorsement, such as 'we adopt the reasoning of'
• Positive: Application or following without criticism, signaled by 'see' or 'accord'
• Neutral: Background context or factual analogy with no evaluative language
• Negative: Criticism, limitation, or distinguishing, often introduced by 'but cf.' or 'however'
• Strongly Negative: Express overruling or rejection, marked by 'overruled' or 'abrogated'

This granularity is essential for authority propagation algorithms that must discount negative treatments to avoid inflating the influence of weakened precedents.

Sentiment classification relies on detecting explicit performative verbs and contrastive discourse markers that signal the citing court's stance. These lexical cues form the feature set for supervised NLP models.

• Positive signals: 'follow,' 'adopt,' 'reaffirm,' 'see generally,' 'accord'
• Negative signals: 'overrule,' 'abrogate,' 'decline to follow,' 'reject,' 'disapprove'
• Contrastive markers: 'but cf.,' 'however,' 'nonetheless,' 'although'
• Limiting language: 'limited to its facts,' 'narrow,' 'distinguish'

Advanced models incorporate syntactic dependency parsing to capture negation scope, ensuring that 'we do not overrule' is correctly classified as positive rather than negative.

Once classified, sentiment polarity is converted into a numerical edge weight that modulates how influence propagates through the citation graph. This transforms a static network into a dynamic authority model.

• Positive edges carry a weight of +1.0, fully transmitting precedential support
• Neutral edges receive a reduced weight of +0.2 to +0.5, acknowledging citation without endorsement
• Negative edges are assigned a weight of -0.5 to -1.0, actively subtracting from authority scores
• Overruling edges may sever the connection entirely, removing the overruled node from active precedent chains

These weighted edges feed into PageRank variants and Graph Neural Networks, enabling the system to compute an authority score that reflects not just citation volume but citation quality.

Citation sentiment is inherently time-dependent. A case may be positively cited for decades before a doctrinal shift triggers a wave of negative treatment. Temporal sentiment analysis tracks this evolution.

• Sentiment drift detection identifies cases where the ratio of positive to negative citations changes significantly over a defined time window
• Pre-shepherdized embeddings capture a case's vector representation before a major overruling event, enabling historical state reconstruction
• Temporal edge attributes store the date and sentiment of each citation, allowing graph queries like 'show all cases that criticized Smith v. Jones before 2020'

This temporal layer is critical for case outcome prediction models, which must use the authority state as it existed at the time of the predicted decision, not the present day.

Citation sentiment directly enhances legal retrieval-augmented generation (RAG) systems by ensuring that retrieved authorities are not only semantically relevant but also legally sound.

• Negative treatment filtering excludes or deprioritizes cases that have been overruled or heavily criticized, preventing AI-generated briefs from citing bad law
• Sentiment-boosted reranking promotes cases with sustained positive treatment from higher courts, aligning retrieval results with precedential hierarchy
• Shepard's signal integration maps commercial citator treatment codes to internal sentiment scores, combining human editorial analysis with automated NLP classification

This fusion of semantic search and citation sentiment produces a hallucination-resistant legal research pipeline where every cited authority carries a verifiable integrity score.

Training robust citation sentiment classifiers requires expert-annotated datasets where legal professionals label the relationship between citing and cited authorities. This annotation process follows strict taxonomies.

• Annotation schemas typically define 5-7 mutually exclusive treatment categories with detailed coding guidelines and edge-case rules
• Inter-annotator agreement is measured using Cohen's Kappa, with a threshold of 0.7 or higher required for dataset acceptance
• Adversarial examples such as 'respectfully disagree' or 'assuming without deciding' are deliberately included to test classifier robustness
• Multi-label annotation captures citations that serve multiple rhetorical purposes, such as both distinguishing and criticizing a precedent

High-quality ground truth data is the bottleneck for sentiment classification accuracy, and leading legal AI systems invest heavily in domain-expert annotation pipelines.

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. Key properties include:

• Directionality: Edges point from citing case to cited authority
• Weighting: Edges can be weighted by citation sentiment, frequency, or treatment type
• Temporality: Nodes are timestamped, enforcing a natural acyclic flow from older to newer decisions

A graph algorithm that iteratively distributes precedential influence scores across a citation network, often using PageRank variants. The core principle is recursive: a case is authoritative if it is cited by other authoritative cases. This technique identifies the most legally significant nodes beyond simple citation counts by accounting for the quality and weight of citing sources.

An NLP task that automatically categorizes how a citing case legally treats a cited authority. Standard labels include:

• Overruled: Explicitly invalidated
• Distinguished: Found factually inapplicable
• Followed: Applied as controlling precedent
• Criticized: Questioned but not overruled This classification is critical for computing accurate citation sentiment and authority scores.

A sequential path through a citation graph tracing the logical lineage of a legal principle from its seminal case through subsequent applying, interpreting, and modifying decisions. Traversal algorithms can identify the complete doctrinal ancestry of a legal rule, enabling systems to detect breaks in authority when an intermediate case in the chain receives negative treatment.

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 textual features and its citation neighborhood structure. This enables downstream tasks like case outcome prediction and link prediction by representing each authority as a dense vector informed by its relational context.

The study of citation patterns over time to model how legal authority evolves, ages, or gains influence. By incorporating timestamps into graph models, systems can detect trends such as:

• Precedent aging: Declining citation frequency over time
• Citation cascades: Bursts of citations following a landmark decision
• Doctrinal shifts: Changes in treatment sentiment across temporal windows