💬 NLP

MiNER: A Two-Stage Pipeline for Metadata Extraction from Municipal Meeting Minutes

This paper introduces MiNER, a two-stage pipeline combining question answering and Transformer-based models for extracting metadata from heterogeneous municipal meeting minutes, which demonstrates strong in-domain performance and establishes the first benchmark for this task despite challenges in cross-municipality generalization.

Original authors: Rodrigo Batista, Luís Filipe Cunha, Purificação Silvano, Nuno Guimarães, Alípio Jorge, Evelin Amorim, Ricardo Campos

Published 2026-03-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Rodrigo Batista, Luís Filipe Cunha, Purificação Silvano, Nuno Guimarães, Alípio Jorge, Evelin Amorim, Ricardo Campos

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of a crime scene, your crime scene is a massive, messy pile of city council meeting notes.

These notes are the official records of how a city makes decisions. They contain gold: Who was there? When did they meet? What was the meeting about? When did it start and end? But here's the problem: these notes are written by humans, for humans. They are messy, inconsistent, and look different in every city. One city might put the date at the top; another might hide it in the second paragraph.

Trying to find this specific information automatically is like trying to find a needle in a haystack where the haystack is made of different types of hay, and the needles are sometimes made of rubber.

This paper introduces MiNER (Minutes Named Entity Recognizer), a smart robot detective designed specifically to solve this messy problem. Here is how it works, broken down into simple steps:

The Problem: The "One-Size-Fits-All" Failure

The authors tried using standard "AI detectives" (general-purpose models) to find this info. But these general AI models are like a Swiss Army knife: great for opening bottles and cutting string, but terrible at performing surgery. They are trained on general text (like news articles) and don't understand the weird, specific jargon of city council meetings. They get confused and miss the details.

The Solution: The Two-Stage Pipeline

Instead of trying to read the entire 50-page document at once, MiNER uses a two-stage strategy. Think of it as a two-person team: The Scout and The Surgeon.

Stage 1: The Scout (Finding the Clues)

The first job is to find where the important info is hiding.

The Analogy: Imagine you are looking for a specific recipe in a giant cookbook. Instead of reading every single word from page 1 to page 500, you first scan the Table of Contents or the headers to find the "Start" and "End" of the recipe section.
What MiNER does: It uses a "Question Answering" model (The Scout) to scan the document and say, "Okay, the meeting details are in the first 3 paragraphs and the last 2 paragraphs. Ignore the middle part where they discuss the budget for 20 years."
Why this helps: It cuts out 90% of the "noise" (the boring, irrelevant text) so the next step doesn't have to waste time on it.

Stage 2: The Surgeon (Extracting the Details)

Once the Scout has found the relevant sections, the Surgeon steps in.

The Analogy: Now that you have the specific recipe page, you need to pull out the exact ingredients: "2 cups of flour," "1 egg," "300 degrees."
What MiNER does: It uses a specialized AI (The Surgeon) to look only at those specific paragraphs and pull out the exact metadata: Meeting Number, Date, Location, Who attended, Start Time, End Time.
The Secret Sauce (Deslexicalization): To make sure this Surgeon works for any city (not just the one it was trained on), the team taught it to ignore specific names. If the text says "Meeting in Porto," the AI is trained to see it as "Meeting in [CITY]." If it says "Mayor Silva," it sees "Mayor [NAME]." This is like teaching a detective to recognize a "suspect" without caring if the suspect is named Bob or Alice. This helps the AI generalize to new cities it has never seen before.

The Results: The Specialist vs. The Generalist

The team tested MiNER against two famous, powerful AI giants (Gemini and Phi), which are like "Super-Computers" that know everything about the world.

The Result: MiNER (the specialist) crushed the Super-Computers.
- Accuracy: MiNER was much better at finding the right info.
- Speed: MiNER was 1,800 times faster. The Super-Computers took over 12 minutes to process a document; MiNER took less than a second.
- Cost & Energy: MiNER used almost 400 times less energy (carbon footprint). Using the Super-Computers was like driving a massive semi-truck to deliver a single letter; MiNER was like riding a bicycle.

The Catch: The "New City" Problem

There is one hiccup. When the team tested MiNER on a city it had never seen before (a "cross-city" test), its performance dropped a bit.

The Analogy: It's like a detective who is an expert in New York City. If you send them to Tokyo, they know the concept of a city, but they get confused by the specific street names and local slang.
The Fix: The paper shows that if you just show the AI one or two examples from the new city, it learns instantly and gets back to near-perfect performance.

The Bottom Line

This paper gives us a new, highly efficient tool called MiNER. It proves that for specific, messy jobs like reading city council minutes, you don't need a giant, expensive, slow "Super-Computer." You need a specialized, two-step team that knows exactly where to look and how to ignore the noise.

It's the difference between hiring a generalist who tries to do everything and hiring a master locksmith who can open a specific door in seconds.

1. Problem Statement

Municipal meeting minutes are critical official records of local governance, containing decisions, participants, and voting outcomes. However, they present significant challenges for Information Retrieval (IR) and automated processing due to:

Heterogeneity: They exhibit highly variable formats, writing styles, and linguistic structures across different municipalities.
Lack of Standardization: Unlike parliamentary datasets, metadata (e.g., meeting number, date, location, participants, session times) is rarely standardized or explicitly structured.
Domain Specificity: General-purpose Named Entity Recognition (NER) models fail to capture these fine-grained, domain-specific categories.
Locating Metadata: Existing approaches often assume metadata sections are pre-identifiable, ignoring the challenge of automatically locating the specific text segments containing this information within a raw document.

2. Methodology: The MiNER Pipeline

The authors propose MiNER, a two-stage pipeline designed to transform raw documents into structured metadata entities.

Stage 1: Metadata Boundary Detection (MBD)

Goal: To identify the specific "opening" and "closing" text segments of a document where metadata is concentrated, thereby reducing noise.
Approach: An extractive Question Answering (QA) model is employed.
- Two prompts are used per document: one to find the start/end of the opening segment and one for the closing segment.
- The model outputs the sentence indices $(\hat{a}_t, \hat{b}_t)$ defining the span.
- If no metadata is found, it predicts a null segment.
Output: A reduced text region $R$ (concatenation of opening and closing segments) passed to Stage 2.

Stage 2: Metadata Entity Recognition (MER)

Goal: To perform fine-grained entity extraction and classification on the reduced text $R$ .
Models: Transformer-based models (BERTimbau for Portuguese, XLM-RoBERTa for English), trained with standard token-level classification (BIO tags) and optionally a CRF (Conditional Random Field) layer.
Deslexicalization Strategy: To improve generalization across different municipalities, a deslexicalization function $\phi(\cdot)$ $ϕ (\cdot)$ is applied during training:
- Participants/Locations: Replaced with synthetic values (using the Faker library) with 60% probability.
- Dates/Times: Randomly varied in format or content with 30% probability.
- Municipality Names: Replaced with a placeholder token (@MUNICIPIO).
Pipeline Formula: $F(D) = f_{NER}(f_{BD}(D))$ , where the QA model filters the input for the NER model.

3. Experimental Setup

Dataset: The CitiLink dataset, comprising 120 Portuguese municipal minutes from six municipalities (180 metadata segments total).
- Annotations: 8 categories including meeting number, type, date, location, start/end times, and participants (president, councilors with presence status).
- Privacy: Personal data was anonymized.
- Cross-lingual: An English version was created via automatic translation (Azure) and human correction.
Baselines:
- MBD: BM25 and Dense Retriever (unsupervised).
- MER: Large Language Models (LLMs) including Phi (open-weight) and Gemini (closed-weight).
Evaluation Metrics: Exact Match (EM), F1-score, Precision, Recall, Inference Time, and Carbon Footprint (via Code Carbon).
Generalization Tests: Leave-one-out cross-validation (training on 5 municipalities, testing on 1) and incremental learning (adding one annotated document at a time).

4. Key Results

Performance Comparison

Boundary Detection (MBD): The transformer-based QA model (XLM-RoBERTa) significantly outperformed unsupervised baselines (BM25, Dense Retriever), achieving an F1-score of 0.826 (Portuguese) vs. ~0.11 for baselines.
Entity Recognition (MER):
- Specialized models achieved state-of-the-art results: BERTimbau + Deslex reached 0.96 F1 (Portuguese) and XLM-RoBERTa + Deslex reached 0.94 F1 (English).
- Core metadata (dates, times, meeting numbers) were extracted with near-perfect accuracy (F1 = 1.00).
- LLM Performance: General-purpose LLMs performed poorly. Gemini achieved only 0.27 F1 (high precision, very low recall) due to fragmented outputs. Phi failed to produce reliable structured JSON.
Efficiency: The MiNER pipeline was drastically more efficient than LLMs:
- Inference Time: ~0.4s vs. 737s (Gemini).
- Carbon Footprint: ~400x lower emissions.

Generalization and Robustness

Cross-Municipality: Performance dropped in leave-one-out settings (F1 ~0.80 for PT, ~0.72 for EN), indicating that stylistic variability between municipalities hinders transfer learning.
Deslexicalization Impact: While intended to help, deslexicalization slightly reduced performance in leave-one-out tests, primarily increasing boundary errors and reducing recall in Portuguese.
Incremental Learning: The model adapted rapidly; adding just 1–3 annotated documents per municipality raised F1 scores above 0.95, demonstrating the need for minimal local supervision.

Ablation Study

Removing Stage 1 (MBD) and applying NER to the full document resulted in a minor F1 drop (0.965 $\to$ 0.945) but increased training time from 1 minute to 50 minutes, proving the value of the segmentation stage for efficiency and noise reduction.

5. Key Contributions

Novel Pipeline: Introduction of a two-stage pipeline combining QA-based boundary detection with transformer-based NER, specifically tailored for the unstructured nature of municipal minutes.
Deslexicalization: Implementation of a domain-specific deslexicalization strategy to handle entity variability across municipalities.
Benchmarking: Establishment of the first benchmark for this domain, including the CitiLink dataset, fine-tuned models, and comparative analysis against state-of-the-art LLMs.
Efficiency Demonstration: Proof that specialized, smaller models outperform massive general-purpose LLMs in both accuracy and computational cost for structured extraction tasks.

6. Significance and Future Work

Significance: This work addresses a critical gap in local governance digitization. It demonstrates that for highly structured, domain-specific tasks, specialized pipelines are superior to "one-size-fits-all" LLMs in terms of accuracy, cost, and environmental impact.
Limitations: Generalization across municipalities remains a challenge due to linguistic and structural diversity.
Future Directions: The authors plan to explore adaptive fine-tuning strategies, domain-agnostic representations, and integration with downstream tasks like meeting summarization and document linking.

Conclusion: MiNER provides a robust, computationally efficient foundation for extracting structured metadata from municipal records, outperforming general-purpose LLMs while establishing a new standard for research in this specific domain.