Blog · arXiv Analysis · Published: August 12, 2026 · Modified: August 12, 2026 · Last reviewed: August 12, 2026

The Parliamentary Voice Becomes the Authority Score

ParliamentRAG tries to prevent a parliamentary summary from collapsing into the voices that occur most often. It also computes who counts as a topic authority and chooses one primary expert for each political group.

That makes the ranking formula an editorial institution. Its value is not neutrality, but the possibility of inspecting and contesting how the system distributes attention.

The Paper

The source is Mirko Tritella, Riccardo Pozzi, and Matteo Palmonari’s Who Speaks Matters: Authority-Aware Multi-View RAG over Italian Parliamentary Proceedings, arXiv:2608.13410v1 [cs.AI], submitted August 13, 2026. It presents and evaluates ParliamentRAG; it does not establish that generated summaries are politically neutral or that they improve citizen understanding.

This differs from the site’s curated-corpus essay, which asks which sources enter a civic answer. Here the corpus is one official parliamentary record, and the contested choice is which speakers within that record receive priority.

The Record Becomes a Graph

For its February 2026 snapshot, the paper reports 608 plenary sessions, 6,010 debates, 40,416 speeches, 151,073 retrieval chunks, and ten parliamentary groups from the Italian Chamber of Deputies’ XIX Legislature. The graph also connects speakers to group and committee memberships, institutional roles, legislative acts, profession, and education; character offsets preserve a path from cleaned chunks to original transcript text.

The retrieval pipeline combines semantic search over speeches with graph search over relevant acts, signatories, and committee membership. It then reranks evidence for relevance, political salience, group coverage, speaker diversity, and authority. A group without retrieved evidence is marked as unsupported instead of receiving invented material.

Authority Is a Policy Choice

The authority score combines six components whose weights were set empirically through expert judgment rather than learned from data: committee membership and speech interventions each receive 0.25, legislative acts 0.20, profession 0.15, education 0.10, and institutional role 0.05. Activity signals decay over time, while biographical and institutional attributes are compared semantically with the query.

Authority has the smallest weight, 0.05, in the overall evidence reranker, which the authors say is meant to keep it from overriding relevance or marginalizing smaller groups. Yet the same score selects the highest-ranked speaker in each group’s evidence pool as that group’s primary expert. The system therefore uses authority gently at one stage and decisively at another.

Group Coverage Is Not Voice Coverage

The generation design creates a section for every parliamentary group and places evidence from its selected expert first. This is a concrete defense against a common-speaker frequency distribution becoming the apparent political landscape. It also makes missing group evidence visible.

But a parliamentary group can contain internal disagreement, changing positions, specialist minorities, and members whose relevant contribution does not resemble the system’s credential-and-activity profile. One representative per group is a tractable summary rule, not proof that the group has one voice. Democratic mandate, topical expertise, institutional seniority, and factual reliability remain different concepts.

What the Evaluation Establishes

The benchmark used 15 policy topics selected from 51 suggestions by 17 Italian citizens interested in politics. Six parliamentary journalists, collaborators, and policy analysts produced 67 blind paired evaluations; only two assessed every topic. NotebookLM received a separate context for each topic containing about 150 relevant and 150 distractor chunks.

In the automated results, quotation-bearing group coverage was 0.97 for ParliamentRAG and 0.95 for NotebookLM; exact quotation faithfulness was 1.00 and 0.95. Overall human satisfaction was 4.24 and 4.27. Source-related ratings descriptively favored ParliamentRAG, while answer quality and clarity favored NotebookLM, but no human-evaluation difference remained statistically significant after multiple-comparison correction. The relevant NotebookLM chunks came from ParliamentRAG’s own retrieval and reranking, so the authors correctly frame this as a curated-context response comparison rather than an independent end-to-end retrieval contest.

Verbatim Is a Boundary, Not a Verdict

ParliamentRAG does not ask the language model to compose quotations. It inserts placeholders that are replaced deterministically from source offsets, producing the reported perfect exact-match score. That is a meaningful architectural guarantee against altered quotation text. It does not establish that a true quotation is representative, sufficiently contextualized, or the best evidence for a group’s position. String fidelity and editorial adequacy need separate tests.

The Evidence Boundary

The authors’ limitations identify the 15-topic benchmark, moderate statistical power, and lack of internal ablations. The experiment covers one chamber, one legislature, one snapshot, and one structured question form. It does not test citizen comprehension, corrections requested by quoted speakers, contested authority labels, long-term use, or whether summaries alter political judgments. The released code and configuration make the chosen weights inspectable, but inspection is not external validation of what authority ought to mean.

The Authority-Score Receipt

A governed parliamentary answer should record the corpus and legislature snapshot, query and rewrite, committee mapping, candidates from both retrieval channels, relevance and coverage scores, complete authority formula, per-speaker component values, time-decay rule, group affiliation date, chosen expert and rejected alternatives, missing-group evidence, transcript offsets and surrounding context, generation model and prompt, metric definitions, reviewer, correction path, and code commit.

The Spiralist boundary is that access to a democratic record must not hide a new gatekeeper inside the ranking layer. The system may order evidence, but the public should be able to see which definition of authority performed the ordering and reopen the choice.

Sources

Return to Blog