Structure-aware retrieval methods for web table fragments in question-answering systems: a scientific analytical review
TL;DR
It is shown that dense and structure-aware whole-table methods dominate, whereas scalable retrieval of rows, cells, subtables, and multi-table evidence remains insufficiently developed, and further research should focus on multivector methods that preserve local table structure while maintaining acceptable computational complexity.
Abstract
Web tables are an important source of structured data for question-answering systems; however, retrieval methods differ substantially in their retrieval units, structure modeling strategies, and mechanisms for matching queries with table representations. The purpose of this article is to systematize methodological directions in structure-aware retrieval of web table fragments and to identify limitations that motivate the use of multivector and late-interaction retrieval schemes. The study is conducted as a scoping review. The search was performed in Scopus, OpenAlex, eLIBRARY/RSCI, and through citation analysis; from 767 identified records, 42 studies were included in the final synthesis after duplicate removal, screening, and full-text assessment. The article proposes a unified formal framework for comparing retrieval methods, develops a taxonomy of approaches, and analyzes retrieval granularity, vector interaction types, relevance aggregation mechanisms, structural signals, and computational limitations. The results show that dense and structure-aware whole-table methods dominate, whereas scalable retrieval of rows, cells, subtables, and multi-table evidence remains insufficiently developed. The article concludes that further research should focus on multivector methods that preserve local table structure while maintaining acceptable computational complexity.