ETD

• BM25
• Dense retrieval
• Document chunking
• Information retrieval
• Long-document retrieval
• Multi-granularity retrieval

Long-document retrieval systems divide each source into chunks: text units that are independently represented, indexed, and retrieved. Granularity controls a fundamental trade-off. Fine chunks isolate local evidence but can lose needed context; coarse chunks preserve context but may dilute query-specific signals.
Because the useful scale depends on document structure and the information need, a single-granularity index imposes the same compromise on every document and query.
Adaptive approaches address this limitation with a router, planner, or specialised encoder that selects among representation scales. This adds a decision mechanism whose training, errors, and computational requirements become part of the retrieval pipeline. Their gains therefore combine the value of multiple granularities with the quality of the selector, rather than isolating the contribution of the candidate representations themselves.
This thesis asks whether useful granularities can instead emerge directly from retrieval. Fine, medium, and coarse chunks are generated independently and pooled in one candidate space. A standard retriever scores all chunks together, allowing relevance scores to determine which scales enter the ranking without routing, quotas, granularity-specific fusion, or a model trained to choose among them.
The method is tested against matched single-granularity controls across four long-document question-answering collections, four chunking families, three dense retrievers, and BM25. Dense retrieval shows a broad advantage across all three models, all four chunking families, and the fine, medium, and coarse controls. BM25 shows a less uniform trade-off, with outcomes varying across datasets and chunking approaches.
Overall, the results show that direct competition among granularities is effective without an explicit selector in dense long-document retrieval; the same design benefits BM25 only in some datasets and against some single-granularity controls.