Source-linked AI summary
Who Speaks Matters: Authority-Aware Multi-View RAG over Italian Parliamentary Proceedings
Mirko Tritella, Riccardo Pozzi, Matteo Palmonari
TL;DR
Parliamentary proceedings are voluminous and fragmented, while naive RAG can distort representation, overlook topical expertise, and misattribute citations. ParliamentRAG uses graph-based multi-view retrieval and query-dependent authority scoring, achieving stronger source-oriented results than NotebookLM while NotebookLM remains stronger on fluency-oriented dimensions.
Problem
Parliamentary records are voluminous and fragmented, while naive RAG can overrepresent vocal actors, ignore topical expertise, and misattribute citations.
Method
ParliamentRAG combines graph-based multi-view retrieval with query-dependent authority-aware reranking using speaker attributes and parliamentary activity, then generates summaries with transcript quotations.
Results
1.00 vs. 0.95 quotation faithfulness and 0.97 vs. 0.95 group coverage; ParliamentRAG was preferred on source-related dimensions, while NotebookLM scored higher on fluency.
Takeaways & Limitations
ParliamentRAG matches NotebookLM on overall satisfaction while excelling on group coverage, quotation faithfulness, and source authority.
Takeaways & Limitations
The evaluation covers only 15 topics and lacks internal ablation experiments, leaving individual architectural contributions unresolved.
Abstract
from arXiv · showhide
Parliamentary proceedings are a primary record of democratic deliberation, yet their volume and fragmentation make multi-perspective access difficult for citizens, journalists, and researchers. Applying Retrieval-Augmented Generation (RAG) to parliamentary transcripts introduces three specific risks: dominance of the most frequent speakers, inability to weight speakers according to topical expertise, and citation misattribution in politically sensitive text. We present ParliamentRAG, a RAG system for the Italian Chamber of Deputies that addresses these risks jointly. Its core contribution is a topic-dependent authority model that estimates each speaker's authority as a function of the current query, combining interpretable components such as profession, education, and previous interventions. Given a user query, the system retrieves relevant speech chunks, identifies topic-relevant experts across parliamentary groups, and generates a summary synthesizing their perspectives, accompanied by supporting quotations. ParliamentRAG is evaluated against Google NotebookLM on 15 policy topics via a two-level protocol combining automated metrics and blind A/B human evaluation by six domain experts. The system achieves higher coverage across political groups (0.97 vs. 0.95), perfect quotation faithfulness (1.00 vs. 0.95), and stronger expert preferences on source-related dimensions, while NotebookLM remains stronger on prose-oriented dimensions.
1 Introduction
ParliamentRAG addresses the scale and fragmentation of Italian parliamentary debates by generating faithful, balanced, multi-view summaries. It combines group coverage, query-dependent authority, and quotation traceability, and outperforms NotebookLM on citation faithfulness and source-related evaluation dimensions.
- Motivation: The XIX Legislature had produced 608 plenary sessions, 6,010 debates, and 40,416 speeches by February 2026, making manual navigation excessively costly.Keyword-based institutional portals return isolated documents rather than synthesized multi-perspective answers.
- Motivation: Naive LLM use risks over-representing vocal actors, flattening speakers to content similarity, and hallucinating or misattributing politically sensitive quotations.These failures reflect corpus distributional asymmetry, ignored speaker metadata, and citation reliability problems.
- ParliamentRAG: ParliamentRAG generates faithful, multi-view, balanced summaries of positions expressed by the Italian Chamber’s ten parliamentary groups.The application targets journalists, analysts, and Italian citizens, with Knowledge Graphs supporting its development and data representation.
- ParliamentRAG: Its design enforces multi-view representation, query-dependent authority awareness, and quotation traceability to verbatim official transcripts.Authority is modeled dynamically so the most relevant voices emerge for each query, while quotations remain directly traceable.
2 Related Work
The related work spans five threads covering parliamentary NLP, knowledge graphs, authority modeling, fairness-aware ranking, and faithful attribution. It identifies a gap in integrating these capabilities into a unified RAG architecture, which ParliamentRAG addresses through structured parliamentary graphs, query-dependent authority modeling, and multi-view generation.
- Research Threads: The paper organizes prior research into five threads: parliamentary NLP, parliamentary knowledge graphs, expert and authority modeling, fairness- and diversity-aware ranking, and faithful attribution and grounded generation.
- Parliamentary NLP: Parliamentary NLP resources support stance detection, topic modelling, and speaker profiling, while ideological scaling and embeddings estimate corpus-level political positions without per-query retrieval or multi-view synthesis.
- Knowledge Graphs in the Parliamentary Domain: Parliamentary knowledge graphs improve transparency, interoperability, and legislative analysis, but mainly support structured querying and data integration rather than authority-aware, query-dependent reasoning or multi-view generation.
- Expert Finding, Authority, Multi-View Retrieval, and Fairness: Existing expert-finding and ranking methods provide query-independent authority or diversity objectives, while political debate requires interpretable query-dependent authority and coverage of multiple legitimate viewpoints.No prior work integrates structured parliamentary graphs with query-dependent authority modeling and multi-view generation within a unified RAG architecture.
- Quotation Faithfulness: Prior attribution and verification methods improve quotation quality but permit paraphrase, whereas ParliamentRAG guarantees that every quotation is a verbatim substring of a specific source document.
3 Parliamentary Knowledge Graph
ParliamentRAG uses a unified Italian-language knowledge graph integrating parliamentary proceedings, legislative activity, and speaker metadata from the Italian Chamber of Deputies’ RDF data. A single Neo4j property graph preserves institutional relationships and supports combined semantic, keyword, and structural access.
- Data integration: The knowledge graph integrates plenary and committee transcripts, legislative acts with signatories, deputy biographies, and temporally valid committee memberships.Its source is the Italian Chamber of Deputies’ open RDF data, structured according to the Chamber ontology.
- Unified architecture: A single Neo4j instance serves as the system’s sole data layer, enabling property-bearing relationships, native vector search, and unified graph, keyword, and dense retrieval.Temporal properties such as start_date and end_date are represented on membership relationships, while act-title keyword search uses a full-text index.
- Graph construction: The graph contains 387 deputies, 64 government members, 10 parliamentary groups, 80 committees, 608 sessions, 6,010 debates, 6,515 phases, 40,416 speeches, and 27,576 legislative acts.It contains 232,755 nodes and 488,487 relationships across 13 node labels and 15 relationship types, with proceedings organized as Session → Debate → Phase → Speech → Chunk.
- Context and traceability: The graph links speakers to time-qualified groups, committees, and legislative acts, supporting time-resolved affiliation, authority modelling, retrieval, and exact quotation grounding.Cleaned transcripts remain aligned with raw text, and generated quotations can be linked to source passages through stored character offsets.
4 ParliamentRAG
ParliamentRAG is an interactive web application that converts queries over Italian parliamentary proceedings into multi-perspective, quotation-grounded answers. Its retrieval–ranking–generation pipeline jointly models relevance, political-group coverage, speaker authority, and deterministic quotation faithfulness.
- System interface: The web application supports free-text queries over official parliamentary transcripts and auxiliary tools for legislative-act and speech search, authority analysis, and position visualization.It is built with Next.js and FastAPI and is publicly accessible.
- Retrieval: The pipeline retrieves evidence through dense semantic search and graph traversal, combining thematic relevance with speaker, group, and legislative-activity information.The dual-channel design produces more diverse evidence across political groups and evidence types than either channel alone.
- Retrieval: Committee membership is additionally privileged when a query maps to a relevant parliamentary committee, using institutional affiliation as a proxy for domain expertise.The mapping covers committees such as justice and foreign affairs.
- Ranking and authority: The evidence score weights relevance at 0.35, salience at 0.25, coverage at 0.20, diversity at 0.15, and authority at 0.05.Authority informs selection without determining it, while one highest-authority speaker per group is selected as the group expert.
- Ranking and authority: The query-dependent authority score combines semantic similarity to profession, education, committee, and role attributes with time-decayed legislative-activity and speech-intervention signals.Its empirically assigned weights are interpretable and not learned from data, while current-group affiliation is resolved at query time.
- Generation: The four-stage Analyze–Generate–Integrate–Cite process creates group-specific sections and resolves quotations deterministically from source speeches rather than generating quotation text.Character-offset placeholders are replaced with verbatim source text, preventing quotation hallucinations by construction.
5 Experimental Evaluation
The evaluation compares ParliamentRAG with Google NotebookLM using automated design-principle metrics and blind expert judgments on 15 policy topics. ParliamentRAG leads on group coverage and source-related dimensions, while NotebookLM performs better on prose-oriented quality and clarity.
- Evaluation protocol: The study evaluates both systems through automatic metrics and a blind A/B assessment by six domain experts.The protocol measures coverage balance, quotation faithfulness, analytical adherence, perceived quality, and overall satisfaction.
- Evaluation protocol: The benchmark contains 15 policy topics selected from 51 citizen-survey topics spanning varying polarization and public-discussion intensity.The topics cover eight thematic macro-areas.
- Automated results: ParliamentRAG achieves GQ = 0.97 vs. 0.95 for NotebookLM, with stratified per-group generation providing near-perfect coverage subject to retrieval limits.NotebookLM omits one or more parliamentary groups 5% of the time.
- Human evaluation results: NotebookLM scores higher on Answer Quality (4.30 vs. 4.04) and Answer Clarity (4.51 vs. 4.27), while overall satisfaction is nearly identical at 4.24 vs. 4.27.None of the observed differences is statistically significant after Holm–Bonferroni correction.
- Human evaluation results: ParliamentRAG receives higher ratings for Source Relevance (4.07 vs. 3.84), Source Authority (4.21 vs. 4.00), and Source Coverage (4.64 vs. 4.39).Pairwise preferences likewise favor ParliamentRAG on Source Relevance (33% vs. 19%), Source Authority (30% vs. 13%), and Source Coverage (25% vs. 5%).
6 Discussion
ParliamentRAG outperforms NotebookLM on source-oriented and political-balance dimensions, while NotebookLM is stronger on fluency-oriented aspects. Its quotation faithfulness is structurally enforced, whereas balanced group coverage remains evidence-dependent, and the evaluation is limited by a small benchmark and no ablations.
- Comparative strengths: ParliamentRAG performs better on source authority, source coverage, source relevance, and perceived political balance, while NotebookLM scores higher on answer quality and clarity.The asymmetry reflects ParliamentRAG’s explicit multi-view retrieval and structural guarantees versus NotebookLM’s fluency strengths.
- Architectural guarantees: Quotation faithfulness is the strongest architectural invariant because offset-based transcript retrieval prevents fabricated or altered quotations.Balanced group coverage is a design objective rather than an absolute guarantee because it depends on relevant retrieved evidence being available.
- Limitations: 15 topics and the absence of internal ablation experiments limit the evaluation’s statistical power and leave the contributions of individual architectural components unresolved.The study compares the complete system with a strong external baseline, making component-level attribution a subject for future work.
7 Conclusion
The paper presents ParliamentRAG, an authority-aware, multi-view RAG system for Italian parliamentary proceedings. On a 15-topic benchmark, it matches NotebookLM in overall satisfaction while outperforming on group coverage, quotation faithfulness, and source authority.
- ParliamentRAG is an authority-aware, multi-view RAG system for Italian parliamentary proceedings.
- On a 15-topic benchmark with expert A/B evaluation, ParliamentRAG matches NotebookLM on overall satisfaction and excels on group coverage, quotation faithfulness, and source authority.
- Future work includes personalized ranking, domain-specific embeddings, extension to the Senate and earlier legislatures, and cross-legislature expansion.
Use of Generative AI
Generative AI tools assisted the preparation of the work by refining language and improving the English text’s clarity, fluency, and overall quality.
- Generative AI tools were used during the preparation of the work to assist with language refinement.
- The tools paraphrased selected passages.
- They improved the clarity, fluency, and overall quality of the English text.