💻 computer science

Who Speaks Matters: Authority-Aware Multi-View RAG over Italian Parliamentary Proceedings

This paper introduces ParliamentRAG, an authority-aware Retrieval-Augmented Generation system for Italian parliamentary proceedings that dynamically weights speakers based on query-specific expertise to mitigate dominance biases and citation errors, outperforming Google NotebookLM in political coverage and quotation faithfulness while maintaining strong expert preference.

Original authors: Mirko Tritella, Riccardo Pozzi, Matteo Palmonari

Published 2026-08-14

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Mirko Tritella, Riccardo Pozzi, Matteo Palmonari

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 walking into a massive, noisy library where thousands of people are having a giant, chaotic debate about how to run the world. This isn't just any library; it's the Italian Parliament, a place where elected representatives argue over laws, justice, and the future of the country. The problem is that this library is so huge, and the arguments so scattered across different rooms and days, that it's nearly impossible for a regular person to figure out what everyone actually thinks about a specific topic. If you just ask a super-smart robot to "summarize the debate," the robot might get confused. It might listen only to the loudest voices, ignore the quiet experts, or accidentally invent quotes that never happened. This is the challenge of "Retrieval-Augmented Generation" (RAG), a type of AI that tries to answer questions by reading a bunch of documents first. The big question is: How do we build an AI that doesn't just pick the loudest speaker, but actually understands who is an expert on what, and makes sure it tells the truth without making things up?

This paper introduces a new system called ParliamentRAG, designed specifically to solve these problems for the Italian Chamber of Deputies. Think of ParliamentRAG as a super-organized, fair-minded librarian who doesn't just shout out the first answer they find. Instead, this librarian has a special map (called a Knowledge Graph) that connects every politician to their job, their education, and the laws they've signed. When you ask a question like, "What do the different political groups think about justice reform?", the system doesn't just grab the most common speeches. It first figures out who the actual experts are for that specific topic, ensuring that every political group gets a fair voice, even if they don't speak as often as the others.

The researchers found that their system is incredibly good at two things that other AI tools often mess up. First, it never makes up quotes. If the system says a politician said something, it is 100% guaranteed to be a word-for-word copy from the official record, like a perfect photocopy rather than a shaky memory. Second, it makes sure to include opinions from almost every single political group (97% of them), whereas a popular commercial AI tool they tested (Google NotebookLM) missed some groups about 5% of the time. While the commercial tool wrote slightly smoother and more "polished" sentences, the experts who tested the system preferred ParliamentRAG when it came to finding the right sources and making sure the political balance felt fair. The authors suggest that while AI can be great at writing, it needs this kind of special "authority-aware" structure to be truly trustworthy when dealing with complex political debates.

The Core Idea: Why "Who" Matters as Much as "What"

To understand why ParliamentRAG is special, you have to understand the three traps that usually catch AI when it reads political debates:

The "Loudmouth" Trap: AI often thinks the person who speaks the most is the most important. But in politics, a quiet expert on healthcare might know more than a loud politician who just talks about everything.
The "Flattening" Trap: AI often treats everyone's opinion as equal, ignoring that some people are actual experts on the topic while others are just making small comments.
The "Hallucination" Trap: AI sometimes invents quotes or misattributes them, which is dangerous when you are talking about laws and government actions.

ParliamentRAG fixes this by using a Knowledge Graph. Imagine a giant spiderweb where every node is a person, a law, or a speech, and the strings connecting them show relationships. This web knows that "Senator Rossi" is a member of the "Justice Committee," has a degree in "Law," and signed a specific bill last year. When you ask a question, the system doesn't just look for keywords; it looks at this web to calculate an "Authority Score." It asks: "Is this person an expert on this specific topic right now?" It weighs their education, their job, and their recent activity to decide whose voice should be heard.

How the System Works: A Four-Step Dance

The system operates like a highly efficient newsroom with four distinct steps:

The Dual-Channel Search: When you ask a question, the system uses two search engines at once. One looks for speeches that sound like your question (semantic search). The other looks at the spiderweb (the graph) to find people who have actually done work on that topic, like signing laws or serving on committees. This ensures it finds both the relevant words and the relevant experts.
The Authority Score: Once it has a list of potential speeches, it calculates a score for every speaker. It combines their background (education, profession) with their recent actions (how many laws they signed, how many speeches they gave). Crucially, this score changes depending on the question. A speaker might be an authority on "Immigration" but not on "Economy."
The Multi-View Selection: The system picks the top expert from each of the ten political groups. This guarantees that the final answer isn't just one side of the story; it's a chorus of different perspectives.
The "No-Fluff" Generation: This is the most unique part. When the AI writes the summary, it doesn't write the quotes itself. It writes placeholders like [QUOTE FROM SPEECH #45]. Then, a separate step goes back to the original, official transcript and pastes the exact words from that speech into the slot. This means the system literally cannot lie about what was said.

The Results: Fairness Over Fluency

The researchers tested ParliamentRAG against Google NotebookLM, a powerful commercial AI, using 15 different political topics. They asked six human experts (journalists and policy analysts) to judge the results.

Here is what they found:

Perfect Truthfulness: ParliamentRAG achieved a 1.00 score for "Quotation Faithfulness," meaning every single quote was a perfect match to the original text. NotebookLM got 0.95, meaning about 5% of its quotes were slightly off or paraphrased.
Better Coverage: ParliamentRAG included representatives from 97% of the political groups, while NotebookLM only managed 95%.
Expert Preference: When the human experts were asked which system they preferred, they chose ParliamentRAG more often for "Source Coverage" (25% vs 5% preference) and "Source Authority." They felt ParliamentRAG did a better job of finding the right experts and balancing the different political views.
The Trade-off: NotebookLM did win on "Answer Quality" and "Clarity." Its sentences flowed better and sounded more like a polished essay. However, the experts noted that while NotebookLM sounded nicer, ParliamentRAG was more trustworthy and balanced.

What This Means

The paper suggests that for serious tasks like analyzing government debates, being "nice-sounding" isn't enough. You need a system that is structurally designed to be fair and accurate. The authors argue that you can't just tell an AI to "be fair" with a simple instruction; you have to build the fairness into the system's architecture, like ParliamentRAG did with its authority scores and strict quote-checking.

While the study was limited to 15 topics and didn't test every possible scenario, the results strongly suggest that adding a layer of "authority awareness" and strict source tracking makes AI much more reliable for civic information. The authors conclude that their system matches the performance of top commercial tools on general satisfaction but beats them significantly on the specific dimensions that matter most for democracy: balance, accuracy, and knowing who is actually speaking.

Technical Summary: Authority-Aware Multi-View RAG over Italian Parliamentary Proceedings

1. Problem Statement

Parliamentary proceedings serve as a primary record of democratic deliberation, yet their sheer volume and fragmentation hinder multi-perspective access for citizens, journalists, and researchers. Manual navigation is costly, and standard keyword-based search portals return isolated documents rather than synthesized, multi-view answers.

Applying standard Retrieval-Augmented Generation (RAG) to political transcripts introduces three specific risks identified by the authors:

Dominance of Frequent Speakers: Standard LLMs tend to over-represent the most vocal actors, reflecting corpus distributional asymmetry rather than the actual political landscape.
Lack of Topical Authority: Systems often flatten all speakers based on content similarity, ignoring metadata that distinguishes a topical expert from a tangential commenter.
Citation Misattribution: In politically sensitive contexts, even low rates of hallucinated or misattributed citations are incompatible with journalistic or institutional use.

The challenge is to build a system that provides faithful, multi-view summaries of parliamentary positions while ensuring traceability to original speeches and avoiding the over-representation of isolated interventions.

2. Methodology: ParliamentRAG

The authors present ParliamentRAG, a RAG system designed for the Italian Chamber of Deputies. The system is built upon a unified Knowledge Graph (KG) and follows a retrieval–ranking–generation pipeline that enforces multi-view representation, authority awareness, and quotation traceability as first-class constraints.

2.1 Knowledge Graph Architecture

The system operates on a property graph stored in a single Neo4j instance, integrating data from the Italian Chamber of Deputies' open data infrastructure (RDF format).

Schema: The graph models a hierarchical structure: Session → Debate → Phase → Speech → Chunk.
Entities: It includes 387 deputies, 10 parliamentary groups, 80 committees, 6,010 debates, and 40,416 speeches.
Enrichment: Node properties include biographical data (education, profession) embedded as vectors, and time-qualified relationships (e.g., MEMBER_OF_GROUP with start/end dates) to handle coalition shifts.
Retrieval Units: Speeches are segmented into 151,073 chunks. Each chunk stores text, a dense embedding, and character-level offsets (start_char_raw, end_char_raw) to enable verbatim reconstruction.

2.2 Dual-Channel Retrieval

To ensure both semantic relevance and institutional expertise, the system employs a dual-channel retrieval strategy:

Dense Channel: Performs vector similarity search over speech chunks to capture thematic relevance.
Graph Channel: Retrieves evidence based on legislative activity. It identifies parliamentary acts relevant to the query, traverses to their signatories (primary and co-signatories), and retrieves their associated speech chunks. It also privileges chunks from speakers who are members of relevant committees.

2.3 Authority-Aware Reranking

Retrieved evidence is merged and reranked using a composite score:
$score(e) = w_{rr}r(e) + w_{dd}d(e) + w_{vv}v(e) + w_{a}authority(s_e, q) + w_{\sigma}\sigma(e)$
Where $s_e$ is the speaker. The weights prioritize relevance ($0.35$) and salience ($0.25$), while diversity and coverage ($0.15$ and $0.20$) ensure group balance.

Authority Score Calculation:
Authority is computed after retrieval to inform ranking without overriding topical relevance. It is a query-dependent score aggregating:

Semantic Similarity: Cosine similarity between the query and speaker attributes (profession, education, committee roles).
Activity Signals: Time-decayed counts of legislative acts signed and speech interventions.
Temporal Decay: Recent activity is weighted higher than past prominence.
Coalition Affiliation: Resolved at query time to ensure only current group activities are considered.

2.4 Multi-View Generation Pipeline

The generation process follows a four-stage pipeline:

Analyze: The query is split into atomic claims linked to evidence for each parliamentary group.
Generate: A section is produced for each of the 10 parliamentary groups, ordered by the authority score of the speakers within that group. If no evidence exists for a group, this is explicitly stated.
Integrate: Sections are combined into a coherent narrative.
Cite: Crucially, the LLM never generates quotation text. It inserts placeholders with character offsets, which are deterministically replaced by verbatim substrings from the original transcripts. This guarantees 100% quotation faithfulness by construction.

3. Experimental Evaluation

The system was evaluated against Google NotebookLM (a strong commercial baseline using Gemini 3) on a benchmark of 15 policy topics.

3.1 Protocol

Dataset: 15 topics spanning eight macro-areas (e.g., Justice, Immigration, Economy).
Baseline Setup: NotebookLM was provided with a curated context of ~300 chunks per topic (relevant + distractors) to simulate expert retrieval, as feeding the full corpus was deemed unfeasible.
Evaluation:
- Automated Metrics: Coverage (Groups with Quotation, Completeness), Quotation Faithfulness, and Mean Authority.
- Human Evaluation: Six domain experts (journalists, analysts) performed a blind A/B test on 67 paired evaluations across nine Likert dimensions.

3.2 Results

Automated Metrics:

Quotation Faithfulness: ParliamentRAG achieved 1.00 (perfect), whereas NotebookLM achieved 0.95.
Group Coverage: ParliamentRAG covered 0.97 of groups (vs. 0.95 for NotebookLM).
Authority: ParliamentRAG cited speakers with a slightly higher mean authority score (0.53 vs. 0.52).

Human Evaluation:

Overall Satisfaction: Comparable between systems (4.24 for ParliamentRAG vs. 4.27 for NotebookLM).
Fluency: NotebookLM was preferred on prose-oriented dimensions (Answer Quality, Clarity), likely due to its monolithic generation and stronger backbone model.
Source & Balance: ParliamentRAG significantly outperformed NotebookLM on dimensions directly aligned with its design:
- Source Coverage: 4.64 vs. 4.39 (25% preference vs. 5% for NotebookLM).
- Source Authority: 4.21 vs. 4.00.
- Balance Perception: 4.66 vs. 4.48.

4. Key Contributions and Significance

The paper claims the following contributions and significance:

Novel Architecture: It is the first work to integrate structured parliamentary graphs with query-dependent authority modeling and multi-view generation within a unified RAG architecture.
By-Construction Guarantees: The system provides a structural guarantee of verbatim quotation faithfulness, solving a critical issue for civic information systems where hallucination is unacceptable.
Authority Modeling: It introduces a method to estimate speaker authority dynamically based on the query, combining biographical, institutional, and temporal signals, rather than relying solely on content similarity.
Balanced Representation: The system ensures that summaries represent all parliamentary groups, preventing the dominance of frequent speakers.
Empirical Validation: The evaluation demonstrates that while general-purpose LLMs may excel in fluency, architecturally enforced constraints are superior for source reliability, balance, and traceability in political contexts.

The authors conclude that while fluency limitations can be mitigated by better language models, guarantees regarding verbatim attribution and systematic group coverage require specific architectural support that prompting alone cannot provide. The system is deployed live at parliamentrag.it with source code available under an Apache-2.0 license.