Research & Papers

LightGBM+CRF spreadsheet AI matches GPU transformers with 0.937 F1

A lightweight pipeline rivals GPU-based TUTA Transformer without the heavy compute cost...

Deep Dive

Spreadsheets remain a stubborn challenge for automated data extraction due to messy layouts and inconsistent conventions. In this CIKM 2026 paper, researchers Antoine Gauquier, Ioana Manolescu, and Pierre Senellart propose an efficient two-stage pipeline that separates learning from deterministic post-processing. First, a LightGBM classifier examines 65 structured features per cell to assign roles, while a pairwise Conditional Random Field (CRF) enforces spatial consistency across the grid. Second, a five-stage deterministic procedure converts those predicted cell types into table bounding boxes, avoiding the need for heavy neural detection layers.

Benchmarked on StatSheets, a new multilingual dataset of 737 manually annotated sheets from 14 public data providers, the CRF-LightGBM system reaches a Mean File-Macro F1 score of 0.937 on cell-type classification—just 0.6 percentage points behind the GPU-dependent TUTA Transformer, but with substantially lower computational overhead. For table detection, the deterministic method beats region-based baselines and stays competitive with LLM-driven systems like SpreadsheetLLM. The results show that combining structured prediction with rule-based range extraction can deliver near-transformer accuracy at a fraction of the cost, making scalable spreadsheet understanding more accessible for real-world document processing pipelines.

Key Points
  • CRF-LightGBM achieves 0.937 Mean File-Macro F1 on cell-type classification, only 0.6 points below TUTA Transformer
  • Deterministic 5-stage table detection outperforms region-based baselines, rivaling SpreadsheetLLM
  • StatSheets benchmark: 737 annotated multilingual sheets from 14 providers across multiple formats

Why It Matters

Efficient spreadsheet understanding means accurate data extraction at scale, without expensive GPU infrastructure.

📬 Get the top 10 AI stories daily