NLP-based Page Classification for Efficient LLM Extraction from Brazilian Public Tender Documents

2026-04-12 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Expert, quick

Summary

A two-stage pipeline has been developed to efficiently extract product information from lengthy Brazilian public tender documents (editais de licitação) using Large Language Models (LLMs). This approach addresses the computational expense and accuracy degradation LLMs face with large inputs by combining NLP-based page classification with LLM extraction. Researchers created a new dataset comprising 11,190 annotated pages from 350 documents across five product domains. Experiments compared transformer-based classifiers like BERTimbau and DistilBERT against classical machine learning models utilizing engineered features. XGBoost, when paired with domain-specific features, achieved a 97.75% F1-score, surpassing fine-tuned BERT models by more than 4 percentage points. The complete pipeline reduces LLM input tokens by 64-88% while preserving extraction completeness, facilitating cost-effective document processing at scale.

Key takeaway

For AI Engineers and Research Scientists working with LLMs on large, structured documents, consider implementing a pre-classification stage. This method, demonstrated to reduce LLM input tokens by 64-88% while maintaining completeness, can drastically cut computational costs and improve processing efficiency. Prioritize classical machine learning models like XGBoost with tailored feature engineering for the classification task, as they may outperform fine-tuned transformer models in specific document domains.

Key insights

Combining NLP page classification with LLM extraction significantly reduces token count and cost for document processing.

Principles

• Feature engineering can outperform fine-tuned transformers.
• Pre-classification improves LLM efficiency and accuracy.

Method

A two-stage pipeline classifies document pages using XGBoost with domain-specific features, then applies LLM extraction only to relevant pages, reducing input tokens by 64-88%.

In practice

• Use XGBoost for document page classification.
• Develop domain-specific features for NLP tasks.
• Pre-filter LLM inputs to cut costs.

Topics

• NLP Page Classification
• LLM Information Extraction
• Brazilian Public Tenders
• XGBoost Performance
• Document Processing Pipeline

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Comparing LLM and Fine-Tuned Model Performance on NVDRS Circumstance Extraction with Varying Prompt Complexity — LLMs, guided by a context-aware "Complexity Score" for prompt selection, significantly outperform fine-tuned models on inferentially complex, low-prevalence data extraction.
• How Machines Understand Words: Token Classification in NLP — Token classification is foundational for NLP, enabling machines to interpret language by labeling individual words and phrases.
• OCR vs LLM Extraction — The Evolution of Document AI — OCR digitizes text, while LLMs understand document meaning and extract contextual information.
• Text-as-Signal: Quantitative Semantic Scoring with Embeddings, Logprobs, and Noise Reduction — Text can be transformed into operational, continuous semantic data points for AI engineering via embeddings and logprob scoring.
• Discovery of Legal Patterns in Civil Petitions via LLM-Based Fact Extraction and Density Clustering — LLM-based fact extraction significantly improves legal document clustering by reducing noise and enhancing semantic coherence.
• McKenzie Marshall: NLP in Asset Management (Barings) — Effective NLP in asset management augments human analysis by precisely identifying and contextualizing company-specific information from diverse text sources.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.