Legal search improves when chunks preserve section structure

When a legal search system retrieves the wrong or incomplete provision, a later answer generator has less useful material to work with. The paper asks whether search can do better when its text chunks preserve where each provision belongs and include context from provisions it cites.

StrucChunk keeps legal sections and their context together in searchable chunks, then tests whether this improves retrieval across legal corpora.

The authors compare StrucChunk with recursive splitting on 501 queries about India’s Criminal Procedure Code. Both use the same hybrid search, combining keyword and semantic rankings.

Recall@5, the share of queries with a relevant section in the first five results, rises from 0.557 to 0.804 with StrucChunk.

A separate four-way comparison finds that changing the chunks accounts for 88% of the total gain from recursive chunks with dense search to StrucChunk with hybrid search.

This supports a role for section structure in the CrPC result, while a Thai legal test shows that StrucChunk does not improve on a hierarchy-aware method overall.

These evaluations measure retrieval of relevant legal sections, not whether a system’s final legal answers are correct. Results concern the structured documents and query sets tested; they do not establish performance on unstructured legal documents.