Does Statutory Structure Help? Retrieving Every Indispensable Provision for Sri Lankan Conveyancing Scenarios
Abstract
A conveyancing question is answered only when every provision it turns on is in hand: the operative rule, the definition it relies on and the exception that decides the case. We ask whether structure-aware retrieval finds these complete sets better than flat retrieval does, in Sri Lanka, where prior work is scored on questions generated from the Act text. We build a corpus of 52 principal enactments and 60 amending Acts (4,557 sections, 6,346 typed edges for definitions, exceptions, qualifications, cross-references and amendments); a benchmark of 50 questions in 40 fact patterns from Sri Lanka Law College conveyancing examinations, each mapped to the sections indispensable to a complete answer with verbatim provenance; and seven retrievers, from BM25 to a hierarchical system that routes to Acts, retrieves seed sections and expands them along the typed edges. The principal metric, complete indispensable recall, credits a question only when every indispensable section is retrieved. Against proposed labels, this explicit graph shows no complete-recall gain at 20 on the test split: 0.344 for hybrid BM25 and dense fusion versus 0.312 for the structure-aware retriever; the paired matter-level bootstrap interval includes zero. No system completes a question needing four or more sections, and only two of the sections the structure-aware retriever missed were reachable by any typed edge from a retrieved seed: most proposed dependencies lack an explicit edge from a seed. The agent-produced labels are released on Hugging Face as provisional reference labels and await three-attorney validation under the protocol we specify.