PaperlayerView exact revision on arXiv

Source-linked AI summary

The Italian lawmaking process: a unified and structured dataset

Matilde Contestabile, Chiara Ferrara, Giovanni Parrillo, Lorenzo Emer, Andrea Vandin

arXiv:2610.23741v1cs.CL

TL;DR

Italian parliamentary histories are often available as unstructured text, while systematic analysis requires dated, linked procedural events. The paper combines official legislative sources with LLM extraction validated against manually annotated records to build a dataset covering all 5,346 laws enacted in Legislatures IX–XIX. Gemma 4 (31B) achieved 92.73% exact-match agreement with few-shot prompting.

  • Problem

    Preparatory works describe legislative histories in narrative form, so procedural events must be identified, dated, and linked to laws for systematic comparison.

    Source: Background & Summary · p. 2

  • Method

    The authors combine Normattiva and Senate records and use LLM extraction, validated against manually annotated records, to structure legislative events.

    Source: Background & Summary · p. 2, Technical Validation · p. 5

  • Results

    Gemma 4 (31B) achieved 92.73% exact-match scores with few-shot prompting, the highest among evaluated models, and was selected for the 5,346-law corpus.

    Source: Data sources and acquisition · p. 3, Technical Validation · p. 7

  • Takeaways & Limitations

    The dataset provides a structured basis for empirical analysis of lawmaking processes and application of process-mining techniques to recorded events.

    Source: Background & Summary · p. 2

    The dataset includes enacted laws only, excludes proposals that did not become law, and has metadata availability that varies over time.

    Source: Usage Notes · p. 8

Abstract

from arXiv · show

The Italian legislative process is documented in detailed parliamentary records hosted in institutional websites. However, much of this information is available only as unstructured text, limiting its systematic reuse. We applied a controlled, systematic LLM-based extraction procedure, validated against manually annotated records, to convert these texts into ordered and dated events including, e.g., bill introduction, committee examination, plenary examination, and approval. The resulting dataset reconstructs the procedural histories of all 5346 Italian laws enacted between 1985 and 2026, spanning parliamentary terms IX to XIX, for overall 101413 events. Each event records the activity, date, and chamber, with additional procedural details where available. The dataset supports research in the emerging fields of computational legal and legislative studies, potentially combined with process-oriented data science methods to examine procedural paths and legislative timing across parliamentary terms.

Background & Summary

Italian legislative histories are often available as narrative text, although systematic analysis requires procedural actions to be identified, dated, and linked to each law. The paper uses an LLM-based extraction procedure to build a structured dataset of legislative events for analysis of lawmaking processes.

Background & Summary

  • Italy’s symmetric bicameral system requires the Chamber of Deputies and Senate to approve the same text, creating potentially complex legislative trajectories.Earlier research has examined Italian legislation in relation to productivity, coalitions, bicameralism, and temporal trajectories.Source: Background & Summary · pp. 1–2
  • Narrative preparatory works must be converted into consistent, dated, law-linked procedural events to support systematic comparison of legislative histories.Existing resources make parliamentary texts and legislative information accessible at scale, but often leave further extraction and processing necessary.Source: Background & Summary · p. 2
  • The study uses Gemma 4 to extract legislative events from narrative preparatory works and assesses accuracy against manually annotated records.The approach draws on LLM work in specialized-domain classification and information extraction [18].Source: Background & Summary · p. 2
  • The dataset links a laws registry to an event log that retains original preparatory text and ordered events recording activity, date, and chamber.Additional procedural details are recorded where available.Source: Background & Summary · p. 2
  • The structured dataset provides a basis for empirical analysis of lawmaking and application of process-mining techniques to recorded legislative events.The paper situates these opportunities in computational legal and legislative studies.Source: Background & Summary · p. 2

Methods

The methods combine official legislative sources with an LLM extraction pipeline and post-processing to construct consistent event histories. The pipeline uses few-shot examples, structured outputs, and checks for potential irregularities.

Data sources and acquisition

  • Normattiva laws and Senate parliamentary archives supply statutory texts, preparatory works, and parliamentary metadata for the dataset.The Senate archives record activity in both chambers because each must approve the same text for enactment.Source: Data sources and acquisition · pp. 2–3

Event extraction with a large language model

  • An Italian-language prompt extracts legislative events as structured JSON records using a predefined ten-attribute schema; few-shot prompting supplies three manually annotated legislative-history examples spanning different parliamentary procedures.Events are reported chronologically, with distinct dates or procedural stages recorded separately. Attributes are assigned only when supported by textual evidence.Source: Event extraction with a large language… · p. 3
  • Inference uses greedy decoding with temperature 0.0 and seed 42, grammar-constrained JSON, and a 64k-token context window.The context window was configured to accommodate prompts, source texts, and generated outputs without truncation.Source: Event extraction with a large language… · pp. 3–4

Data post-processing

  • Post-processing standardizes extracted records, while rule-based checks flag anomalies without excluding affected laws.Checks cover chronology, missing dates, invalid chamber transitions, and malformed records; flagged cases remain in the dataset.Source: Data post-processing · pp. 4–5

Data Records

The dataset comprises a laws registry and a linked event log containing original preparatory works and ordered procedural histories. Together, these files cover all 5,346 laws enacted during Legislatures IX–XIX.

Data Records

  • The dataset is released under CC BY 4.0, with code and reproducible evaluation pipelines under the MIT License.The complete dataset is publicly available on GitHub.Source: Data Records · p. 5

Laws Registry (laws registry.csv)

Preparatory Work (preparatory work.jsonl)

Technical Validation

The paper validates event extraction against manually annotated legislative histories using eight open-weight LLMs and both zero-shot and few-shot prompts. Gemma 4 (31B) with few-shot prompting achieved the highest exact-match score and was selected for the full corpus.

Technical Validation

  • Validation used 100 laws sampled proportionally across legislatures, with preparatory works manually annotated for 2,031 events.Three independent annotators produced the reference dataset, resolving isolated disagreements through discussion.Source: Technical Validation · p. 5
  • Evaluation checked schema compliance, exact event-set agreement, and semantic-similarity matching between predicted and reference records.The semantic approach uses vector similarity and one-to-one assignment; exact matching requires all ten event attributes to agree.Source: Technical Validation · pp. 6–7
  • 92.73% exact-match score was achieved by Gemma 4 (31B) with few-shot prompting, the highest exact-match result among evaluated models.Qwen 3.5 (27B) followed at 91.29% in the few-shot configuration.Source: Technical Validation · p. 7
  • Few-shot prompting increased Jaccard similarity for all eight models by an average of 18.3 percentage points.Gemma 4 (31B) also ranked first on semantic matching, with micro-F1 of 99.80% and macro-F1 of 99.87% under few-shot prompting.Source: Technical Validation · p. 7

Usage Notes

The dataset is available in standard data formats and retains source text alongside extracted events. Its quality flags and coverage limits should be considered when analyzing its records.

Usage Notes

  • The dataset is available as CSV and JSON Lines files linked by a shared law identifier, and event sequences can be converted for process-mining tools.The listed tools include PM4Py, ProM, and Disco.Source: Usage Notes · p. 7
  • Coverage is limited to enacted laws, metadata availability varies across legislatures, and same-day events have ordered sequences but no exact times.The earliest legislatures have less metadata than later ones, which include additional bill and procedural attributes.Source: Usage Notes · p. 8
  • The series begins with records from 1985 for Legislature IX and covers Legislature XIX only through June 2026.These are incomplete endpoints of the respective legislatures.Source: Usage Notes · p. 9
  • The suspicious flag marks potential inconsistencies in legislative histories; 4,541 of 5,346 laws (84.9%) were unflagged and 805 (15.1%) had at least one anomaly.Manual inspection found most inconsistencies originated in errors or omissions in the original source text; the flag can identify records for inspection or robustness checks.Source: Usage Notes · p. 9

Data Availability

The dataset is openly available in the project’s GitHub repository under the dataset directory, with a CC BY 4.0 license.

Data Availability

Code Availability

The code for constructing and evaluating the dataset is publicly available on GitHub, with execution and reproduction guidance documented in the repository.

Code Availability

  • GitHub provides the code for dataset construction, few-shot LLM inference, post-processing, anomaly detection, and empirical evaluation.Source: Code Availability · p. 9
  • The code is implemented in Python 3.10+ and released under the MIT License.Source: Code Availability · p. 9
  • The repository documents execution instructions, dependencies, evaluation configurations, and reproduction guidelines.Source: Code Availability · p. 9

Related papers

Loading 2610.23741v1…