Spreadsheet question answering improves by smart workbook selection and exact computation
FICO: Find-Then-Compute for Corpus-Level Spreadsheet Question Answering
Information Retrieval
Summary
Answering questions across many spreadsheets is hard because you need to find the right spreadsheet and calculate answers correctly from large tables. The authors created a method called FiCo that first finds the right spreadsheet by comparing summaries and then runs precise database queries on the full data. This approach improves accuracy a lot compared to earlier methods that guessed spreadsheets or used less exact calculations. The method works well on big collections of spreadsheet questions, showing that combining good selection with exact calculations helps a lot.
What this means in practice
- •For business analysts: Automatically find the correct spreadsheet and compute answers to complex questions over multiple company data files.
- •For data integration engineers: Build tools that link many spreadsheets by identifying and querying the right tables for combined data analysis.
- •For financial services teams: Create customer support tools that find the exact financial workbook and generate precise answers from underlying tables.$Commercial implications: Enables building advanced financial question answering products that improve information retrieval and calculation accuracy for client workbooks.
Authors
Sandarsita Guntupalli, Lu He, Kang Li
Abstract
Question answering over spreadsheet collections requires finding the correct workbook and computing over complete tables. We introduce Find-then-Compute (FiCo), which retrieves document summaries, disambiguates similar workbooks, and executes constrained Structured Query Language (SQL) over the selected full table. On DataBench (80 datasets, 1,810 questions), FiCo reaches 76.2% accuracy: 9.9 points above a strong TableRAG-style baseline on the same frozen workbook choices (66.3%) under the tracks' prespecified evaluators, and 63.4 points above prefix RAG (12.8%). On 508 MiMoTable questions, FiCo reaches 79.7%, versus 22.2% for prefix RAG. Giving the strong baseline the gold workbook raises it from 66.3% to 76.3%, exposing a 10.0-point source-selection cost under fixed compute. Despite 95.1% document recall and 98.5% executable SQL, only 81.3% of questions execute on the gold workbook. FiCo's advantage comes from integrating semantic source selection with exact, schema-grounded computation.