RATIO: Retrieval of Authority via Typed-edge Inference and Ontology

Abstract

We present RATIO, a graph-aware framework for legal precedent search that models the citation network directly rather than relying on semantic similarity. RATIO formalises the Binding Horizon Problem — the question of whether a given precedent is valid authority for a proposition — as a temporally-constrained, jurisdiction-masked graph reachability problem.

Introduction

Legal precedent retrieval is fundamentally different from general semantic search. A case is only good authority if it (1) originates from a binding jurisdiction, (2) was decided before the target proposition's temporal horizon, and (3) has not been overruled or distinguished beyond usefulness. Semantic similarity models ignore all three constraints, producing results that look plausible but are legally invalid.

RATIO addresses this by encoding the citation network as a typed, directed, temporal graph. Each edge carries a treatment type (followed, applied, distinguished, overruled, cited) and a timestamp. Retrieval becomes a constrained reachability query over this graph.

The Binding Horizon Problem

We define the Binding Horizon Problem as follows:

Given a query proposition $Q$ and a candidate precedent $C$, determine whether $C$ constitutes binding authority for $Q$ at time $t$. This requires:

1. Jurisdictional constraint: $C$ must originate from a court whose decisions are binding on the target court
2. Temporal constraint: $C$ must have been decided before $t$
3. Treatment constraint: $C$ must not have been overruled or negatively treated by a higher authority before $t$

Formally, we model this as a reachability problem in a directed acyclic graph $G = (V, E, T)$ where $V$ is the set of cases, $E$ is the set of citation edges with treatment types, and $T$ is the temporal ordering function.

Construction of the Nigerian Citation Graph

We constructed the first open citation graph for Nigerian case law, comprising:

• 14,437 Supreme Court and Court of Appeal judgments
• 51,465 typed citation edges across 5 treatment categories
• 22 years of jurisprudence (2002–2024)
• All 36 Nigerian states represented in the court hierarchy

The graph was built from publicly available judgment PDFs using a combination of OCR, structured extraction, and manual validation of treatment-type annotations by Nigerian legal practitioners.

Results

RATIO achieves +184.0% Hits@10 over the MLP baseline on the Nigerian citation retrieval task. The improvement is most pronounced for queries involving:

• Cross-jurisdictional precedent (where semantic similarity fails to recognise hierarchical relationships)
• Older precedents that remain good law (where recency-biased retrievers incorrectly deprioritise them)
• Distinguishable cases (where surface similarity masks legal inapplicability)

Implications

RATIO demonstrates that for high-stakes legal retrieval, graph structure encodes more signal than document embeddings. We release the citation graph and evaluation framework as open data under a Creative Commons licence, and the retrieval framework as Apache 2.0.