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

The Problem: RAG Systems Amplify the Loudest Voices

Retrieval-Augmented Generation (RAG)—a technique that retrieves relevant documents and feeds them to a language model to generate answers—sounds ideal for parliamentary records. But applying it naively to political transcripts creates three concrete risks. First, frequently-speaking politicians dominate the retrieved results, drowning out less vocal but equally informed members. Second, the system can't tell whether a speaker actually knows what they're talking about on a given topic—a legislator with medical training matters more on healthcare than on infrastructure policy. Third, misquoting politicians in sensitive contexts risks distorting the historical record. Existing systems like Google NotebookLM handle these risks poorly in the parliamentary domain.

The authors claim that explicitly modeling speaker authority as topic-dependent—not just counting who spoke loudest—fixes all three problems simultaneously. They built ParliamentRAG to test this on Italian Chamber of Deputies transcripts.

How Authority-Aware Retrieval Works

The system's core innovation is straightforward: instead of treating all speakers equally, it estimates each legislator's "authority" on the current query topic. This authority score combines interpretable signals: professional background (a physician has authority on health policy), educational credentials, frequency of relevant previous interventions, and political group diversity (to ensure multi-perspective coverage).

Given a user query, ParliamentRAG retrieves relevant speech chunks from the transcript database, then identifies which speakers are topical experts across different parliamentary groups. Finally, it generates a summary that synthesizes these expert perspectives with direct quotations tied to specific speakers and dates—ensuring transparency about who said what.

The design explicitly resists the most-frequent-speaker bias that haunts standard RAG. A legislator who spoke once on a topic with deep expertise outranks someone who mentioned it casually dozens of times.

Results: Better Balance, Perfect Quotes

The authors evaluated ParliamentRAG against NotebookLM on 15 policy topics using two levels of assessment: automated metrics plus blind human evaluation by six domain experts (journalists, researchers, and policy analysts familiar with Italian parliament).

ParliamentRAG achieved 0.97 coverage across political groups versus 0.95 for NotebookLM—meaning it more reliably included diverse viewpoints rather than clustering around the majority. More critically, it achieved perfect quotation faithfulness (1.00 vs. 0.95), meaning every quote was accurate to the original text. In the human evaluation, domain experts preferred ParliamentRAG on source-related dimensions—transparency, attribution, and multi-perspective balance. NotebookLM remained stronger on prose quality, but that's a secondary concern when accuracy and representativeness are at stake.

Why This Matters

Parliamentary proceedings are the written record of democratic deliberation. Citizens, journalists, and researchers should be able to query them for balanced, authoritative answers. Standard RAG fails because it optimizes for retrieval statistics, not democratic representation. ParliamentRAG shows that adding a thin layer of domain-aware authority modeling—interpretable, auditable, topic-sensitive—can fix the bias problem without sacrificing accuracy. The perfect quotation faithfulness matters most: in politically sensitive texts, a misquote isn't just an error, it's a distortion of history.