TD-CRESTS: Top-Down Chunk Retrieval Based on Entity, Section, and Topic Selection

Retrieving specific information from extensive scientific documents presents a significant challenge. Existing information retrieval methods often focus on entire documents, even when only a small portion of a document is relevant. Also, achieving a balance between precise retrieval and optimal time complexity remains a persistent challenge. To address these issues, we propose TD-CRESTS (Top-Down Chunk Retrieval based on Entity, Section, and Topic Selection), a term-based document text chunk retrieval method. TD-CRESTS utilizes a hierarchical context architecture, indexing documents according to topics, named entities, sections, and individual text chunks. Key terms from each context guide a top-down search strategy across the index context levels, prioritizing the most relevant contexts based on their overlap with the query. Our method achieves chunk-level F1-measure of \(71\%\) and \(77.14\%\) on the SciREX and DrugSemantics benchmark datasets, respectively. It is able to handle diverse domains and languages with a balance between information retrieval efficiency and effectiveness.