When a citizen asks "can my landlord keep my deposit if the walls were already damaged?", the article that answers the question often shares almost no words with it. Worse, the answer is rarely one article. The article that names the obligation says "subject to the exceptions in Article 1730" or "as defined in Article 2", and those referenced articles share even less vocabulary with the question. Current retrievers score each article by how similar it looks to the query. Those that prune a document graph also prune by query similarity. Under this regime the provisions a legal answer depends on are exactly the ones that get thrown away. We propose to keep a provision because something relevant points to it, not because it resembles the query. Concretely, we parse the explicit cross-references between statutory articles into a directed dependency graph and collapse its cycles into single units so that it becomes acyclic. We then pass relevance from lexically matched "anchor" articles down to their dependencies in a single ordered sweep.
The opportunity is concrete. Legal retrieval is a measured "lexical desert": A Reasoning-Focused Legal Retrieval Benchmark reports TF-IDF cosine similarity between queries and gold statutes of 0.07–0.09, against 0.27 on Natural Questions, and BM25 Recall@10 of 5.03% on Bar Exam QA. LexRAG shows that feeding an LLM irrelevant or incomplete statutes pushes answer quality below the no-retrieval baseline. Retrieving the right article without the definition or exception it depends on is one form of "incomplete". A retriever that fetches the dependency set, in dependency order, attacks both problems at once.
Core Claim
Our central hypothesis is that, under severe vocabulary mismatch, a meaningful share of gold statutory articles are unreachable by any query-similarity scorer yet reachable in one or two cross-reference hops from an article that is lexically retrievable. We expect a pruning rule based on dependency reachability to recover these articles while shrinking context, where similarity-based pruning discards them.
The acyclic constraint is how we make this rule well-defined without destroying information. Statutes cross-reference each other in cycles. For example, a scope article cites an exception article that cites back to the scope. We do not cut these feedback edges, which would silently drop a genuine mutual dependency. Instead we merge each strongly connected component (a maximal set of articles that all reach each other through references) into one super-node. This follows the cycle-breaking motivation of Preferred Reasoning in ABA by Cycle-Breaking and the convergence concerns raised for cyclic argument graphs by Contribution functions for quantitative bipolar argumentation graphs. Merging keeps every reference but turns the graph into a DAG, a directed acyclic graph. On a DAG, relevance propagation is one exact pass in topological order, with no damping factor, no iteration count, and no oscillation. The same order also gives a principled way to lay out the retrieved context for the generator: definitions and exceptions appear next to, and after, the provisions that invoke them.
What is new relative to the closest work is the pruning criterion and the edge set. Finding the Law (G-DSR) runs a graph attention network over the containment tree of Belgian codes (Code → Book → Title → Chapter → Article). It deliberately omits cross-references because they introduce cycles, and it aggregates over all sibling articles without pruning. QABISAR adds query–article bipartite edges and distills the graph model into a bi-encoder, but still has no reference edges and no dependency direction. CuSINeS uses the tree only at training time to pick hard negatives. Hierarchical Document Refinement (LongRefiner) prunes document trees by query similarity, which is the failure mode we target. DocNavRAG follows lateral citation links with an LLM agent, relies on prompt state to avoid loops, and was not tested in lexical-desert domains. Our method is non-agentic, handles cycles by construction, and is evaluated precisely where lexical similarity fails.
Research Program
The build has three parts.
Graph construction. For each corpus we build a heterogeneous graph with article nodes, containment edges from the division headings (BSARD ships these), and typed cross-reference edges extracted from article text with jurisdiction-specific patterns ("article 1730", "本法第X条", "as defined in §…"). A small classifier assigns each reference a type from the surrounding clause: definition, exception, penalty, procedure, or generic. We then compute strongly connected components and condense them.
Relevance propagation. A first-stage retriever gives each article $v$ an anchor score $s(v)$. For BSARD this is a fine-tuned dense bi-encoder, optionally run on queries expanded by LLM legal-reasoning rollouts. We then sweep the condensed DAG in topological order:
$$ r(v) = \max\Big(s(v),\ \max_{u \to v} \gamma_{\tau(u,v)}\, r(u)\Big) $$
Here $u \to v$ means article $u$ references $v$, $\tau(u,v)$ is the reference type, and $\gamma_\tau \in [0,1]$ is a per-type decay tuned on the development split. There are five parameters in total. We use a max rather than a sum on purpose. A sum would inflate heavily cited general-definition articles, reproducing the hub problem that Ponzetto et al. addressed by pruning high-centrality category nodes. Inside a condensed component, all members receive the component's score.
Pruned context assembly. For generation, we keep the top-$k$ articles by $r$ plus their dependency closure up to depth $d$, subject to a token budget. Siblings in the containment tree are added only if they score on their own, which is the soft counterpart of pruning G-DSR's indiscriminate sibling aggregation. The kept articles are serialized in topological order.
Analysis. Before any modeling we will measure, on each benchmark, the fraction of queries whose gold set contains an article that is (a) in the bottom half of first-stage similarity and (b) reachable within two reference hops from another gold or top-50 article. If this fraction is negligible, the proposal is dead and we will say so. If it is substantial, it defines the evaluation slice where the method must win.
Baselines
Naive reference. BM25 over flat articles, and the same fine-tuned dense bi-encoder that serves as our anchor scorer, with no graph at all. On BSARD, A Statutory Article Retrieval Dataset in French reports Recall@100 of 74.78% for a fine-tuned two-tower CamemBERT and 51.33% for BM25. Any gain we report must be over our own rerun of this bi-encoder, because propagation starts from it.
Strongest structural baseline. QABISAR, which reports R@500 of 43.1 on BSARD against 34.8 for the next best graph baseline (BE+GE-Stat.). These numbers come from a different evaluation protocol than the 74.78% R@100 figure, which is why we will rerun QABISAR and G-DSR under one protocol, with the same encoder backbone, the same training data, and an equal hyperparameter search budget. For the reruns we will use released code if it is available and otherwise reimplement from the papers.
Soft-pruning baseline, the most direct threat. G-DSR's GATv2 encoder run over the containment tree plus our extracted cross-reference edges, cycles included. This tests whether learned attention, a form of soft pruning, already exploits references without any acyclic structure.
Graph-walk baseline. Personalized PageRank from the anchor scores over the same directed reference graph.
Ablations. These isolate the key ingredient. We replace SCC condensation with cutting feedback edges found by depth-first search. We replace dependency propagation with LongRefiner-style similarity pruning at the same token budget. We replace max with sum. We use untyped edges, meaning a single $\gamma$.
What counts as winning. We beat the structural baselines if we achieve higher Recall@100 and MAP on the full test set with non-overlapping bootstrap confidence intervals, and a clearly larger gain on the dependency slice. On the generation side, we win if, at matched context tokens, our pruned context scores above LexRAG's reported pipelines, whose best retriever reaches Recall@10 of 33.33%, and never falls below the zero-shot answer score.
Why not just…?
Why not just expand the query with an LLM, as the reasoning-rollout method does? A Reasoning-Focused Legal Retrieval Benchmark shows rollouts raise BM25 recall by over 6 points and E5-large by nearly 9 points, so this is a real competitor. However, a rollout hypothesizes the principle at issue. It has no way to name the specific cross-referenced exception or definition, which lies one reference hop from that principle in the corpus. We treat expansion as complementary. We will run it as a baseline and also as our anchor scorer. The decisive question is whether propagation still adds recall on the dependency slice when it sits on top of expanded queries. If the gain vanishes there, expansion suffices and we will report that.
Why not just add cross-reference edges to a graph attention network and let attention learn what to ignore? This is the G-DSR-plus-references baseline. It is the likeliest way for us to lose, because attention is soft pruning. Our expectation that it falls short rests on two points. First, message passing mixes features over neighborhoods regardless of edge direction semantics. Second, the lexically distant dependency has a poor embedding to begin with, so attention scores computed from query–node similarity will down-weight it. If this baseline matches us on the dependency slice, our explicit acyclic propagation offers no retrieval benefit. The contribution would then shrink to context ordering and interpretability, and the experiments will show which.
Why not just run personalized PageRank on the cyclic graph, which converges anyway thanks to damping? Convergence is not the issue. The issue is that PageRank accumulates mass in reference cycles and in heavily cited hubs, and it yields no ordering for context assembly. The feedback-edge-cutting and sum ablations test the same failure modes inside our own method. If PageRank ties with us on recall, the acyclic constraint earns its place only through the generation-ordering experiment below. If it ties there too, the constraint is unnecessary.
Why not just use a stronger cross-encoder reranker? A reranker only reorders what the first stage retrieved. It still judges each article against the query in isolation, so it cannot promote an exception article whose text never mentions the user's situation. We will include a cross-encoder reranker over the top 100 as a baseline and report its recall on the dependency slice.
Proposed Experiments
Retrieval benchmarks. We use three: BSARD (22,633 articles, 1,108 queries, with heading annotations), Housing Statute QA (about 6,800 yes/no questions over 2 million state statutes, where cross-references must be extracted at scale), and LexRAG's 17,228 Chinese articles, via the LexiT toolkit. We report Recall@10/100/500 and MAP overall and on the dependency slice.
Generation. For LexRAG we use the toolkit's LLM-as-judge scoring. For Housing Statute QA we use answer accuracy. A Reasoning-Focused Legal Retrieval Benchmark reports GPT-4o-mini accuracy rising from 49.7% to 71.7% with retrieved context, and this is the reference point. We compare our topologically ordered, dependency-closed context against similarity-pruned and unpruned contexts at matched token budgets. We also include a shuffled-order control, which isolates whether ordering itself matters, echoing DiagChain's finding that models retrieve over 80% of the evidence yet fail at ordering.
Decisive checks. Three results would decide the idea. The first is the pre-modeling slice statistic: a negligible fraction of reference-reachable, low-similarity gold articles ends the project. The second is the comparison against G-DSR-plus-references on that slice. The third is the ablation pitting SCC condensation against edge cutting and against PageRank, which tells us whether the acyclic constraint does real work or is only a convenient bookkeeping device.
References
- A Reasoning-Focused Legal Retrieval Benchmark
- LexRAG: Benchmarking Retrieval-Augmented Generation in Multi-Turn Legal Consultation Conversation
- Preferred Reasoning in ABA by Cycle-Breaking
- Contribution functions for quantitative bipolar argumentation graphs: A principle-based analysis
- Finding the Law: Enhancing Statutory Article Retrieval via Graph Neural Networks
- QABISAR: Query-Article Bipartite Interactions for Statutory Article Retrieval
- CuSINeS: Curriculum-driven Structure Induced Negative Sampling for Statutory Article Retrieval
- Hierarchical Document Refinement for Long-context Retrieval-augmented Generation
- DocNavRAG: Document-Structured Graph RAG with Stateful Evidence Construction for Complex Document Question Answering
- Deriving a Large-Scale Taxonomy from Wikipedia
- A Statutory Article Retrieval Dataset in French
- DiagChain: A Diagnostic Benchmark for Evaluating LLM Agents on Evidence-Grounded Attack Chain Reconstruction