Reproduce “SpreadsheetCoder: Formula Prediction from Semi-structured Context” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinyun Chen, Petros Maniatis, Rishabh Singh, Charles Sutton, Hanjun Dai, Max Lin, Denny Zhou
SPREADSHEETCODER is a neural system that automatically predicts spreadsheet formulas by leveraging the rich semi-structured information within tabular contexts, including cell values, headers, and the two-dimensional layout. The system achieves 42.51% top-1 full formula accuracy on a Google Sheets dataset and assists 82% more users in formula composition compared to existing rule-based systems.
Key insights
- Header information is critical for semantic understanding, with its removal causing formula accuracy to drop from 42.51% to 20.47%.
- Combining both row-based and column-based BERT encoders along with convolution layers significantly improves performance, indicating the value of modeling diverse and long-range tabular dependencies.
- Pre-training BERT encoders with English weights yields approximately 10 percentage point improvements, showing that natural language understanding capabilities transfer to spreadsheet header and text interpretation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.