CitiLink-Summ: Summarization of Discussion Subjects in European Portuguese Municipal Meeting Minutes

1. Problem Statement

Municipal meeting minutes are critical for government transparency but present significant accessibility challenges for citizens due to their length, dense formal language, and complex segmentation into multiple discussion subjects. While Automatic Text Summarization (ATS) offers a solution, its application to this domain faces two major hurdles:

• Domain Complexity: Minutes require models to detect key topics, interpret intent, and filter irrelevant context within a highly specific administrative framework.
• Resource Scarcity: There is a severe lack of high-quality, manually curated datasets for European Portuguese, a low-resource language. Existing meeting summarization datasets (e.g., MeetingBank, ICSI) are primarily in English or focus on spoken transcripts rather than formal written minutes. Furthermore, no existing benchmark exists for summarizing specific discussion subjects within Portuguese municipal documents.

2. Methodology

Dataset Construction (CitiLink-Summ)

The authors created a new corpus specifically designed for segment-level summarization:

• Source Material: 120 municipal meeting minutes collected from six Portuguese municipalities (Alandroal, Campo Maior, Covilhã, Fundão, Guimarães, and Porto) covering the 2021–2024 administrative term.
• Annotation Process:
  • Segmentation: Documents were manually segmented into distinct discussion subjects.
  • Summarization: A team of four PhD-level linguistics annotators, supervised by two expert linguists, manually wrote summaries for each segment.
  • Workflow: The process followed a rigorous three-phase cycle (pre-writing, writing, evaluation) with strict guidelines. Annotators defined the theme, identified the main idea, drafted a summary, and revised it based on five parameters: relevance, cohesion, coherence, length, and abstraction.
  • Privacy: All sensitive data was manually anonymized.
• Dataset Statistics:
  • Total Segments: 2,880 discussion subjects.
  • Splits: Randomly split by municipality to prevent data leakage: 60% Training (1,799), 20% Validation (552), 20% Test (529).
  • Abstraction Level: Analysis using Coverage and Density metrics revealed that the summaries are largely abstractive (low density, medium-to-high coverage), requiring significant reformulation rather than simple extraction.

Benchmark Evaluation Pipeline

To establish baselines, the authors implemented a three-stage pipeline:

1. Chunking: Due to context window limitations (256–2048 tokens) in many models, discussion subjects were hierarchically split using a sliding-window approach, summarized individually, and aggregated.
2. Model Selection & Fine-tuning:
   • Encoder-Decoder Models: BART, BART Large, PTT5 (Portuguese T5), LED (Longformer-Encoder-Decoder), and PRIMERA were fine-tuned on the dataset.
   • Large Language Models (LLMs): Qwen2.5-1.5B (open-source) and Gemini-2.5-flash (closed-source) were evaluated using few-shot prompting.
   • Input Constraint: Models were evaluated using only the discussion subject text, excluding the theme metadata used during human annotation to reflect real-world deployment constraints.
3. Evaluation Metrics:
   • Lexical: ROUGE (R-1, R-2, R-L), BLEU, and METEOR (accounting for stemming/synonymy).
   • Semantic: BERTScore (F1, Precision, Recall) to evaluate contextual similarity.

3. Key Contributions

1. CitiLink-Summ Dataset: The first publicly available corpus of European Portuguese municipal meeting minutes with 2,880 manually hand-written, abstractive summaries of discussion subjects.
2. Comprehensive Benchmark: Establishment of baseline performance for both traditional fine-tuned encoder-decoder models and modern LLMs in this specific domain.
3. Resource Release: Public availability of the full corpus, summarization guidelines, dataset samples, and reproducibility code via GitHub.
4. Quality Control Framework: A detailed annotation protocol that ensures high-quality, abstract summaries, addressing the difficulty of generating non-extractive summaries in low-resource languages.

4. Results

The evaluation (Table 1 in the paper) highlights the following findings:

• Model Performance: Larger models consistently outperformed smaller ones.
  • Top Performer: PRIMERA achieved the highest scores across most metrics (ROUGE-1: 66.17, BERTScore F1: 85.79).
  • Strong Contenders: BART Large (ROUGE-1: 68.96) and Gemini-2.5-flash (ROUGE-1: 64.16) also performed well.
  • Lower Performers: Smaller models like Qwen2.5-1.5B struggled significantly (ROUGE-1: 44.24), indicating that parameter size and domain adaptation are critical.
• Metric Consistency: Models showed consistent behavior across lexical and semantic metrics, suggesting they capture both surface-level overlap and deeper semantic meaning.
• Room for Improvement: While models produced coherent summaries, absolute scores remain moderate, indicating that the task is challenging and current state-of-the-art models have not yet mastered the specific nuances of Portuguese administrative text.

5. Significance and Future Work

• Advancing NLP in Low-Resource Languages: This work fills a critical gap in NLP resources for European Portuguese, specifically for administrative and legal domains.
• Civic Technology: The dataset enables the development of tools that improve public access to government decisions, fostering transparency and accountability.
• Foundation for Future Research: The benchmark allows researchers to compare new architectures and techniques.
• Future Directions: The authors plan to expand the corpus, incorporate human evaluations for clarity/accuracy, explore leveraging theme information as input, and release an English version to facilitate cross-lingual research.

In conclusion, CitiLink-Summ provides a vital resource for the NLP community to tackle the complex challenge of summarizing formal administrative texts in a low-resource language, setting a new standard for municipal meeting summarization research.