CivicParse: A Benchmark and Pipeline for Structured Online Deliberation
Abstract
Online deliberation platforms promise scalable collective intelligence, yet their free form threads are difficult to navigate, summarize, and moderate. We argue that progress requires treating structured deliberation as a formal natural language processing (NLP) problem with civic significance: reliably mapping raw discussions into a deliberation native schema so that key barriers, solutions, metrics, and stances are visible at scale. We introduce CIVICPARSE, a two stage pipeline that operationalizes this problem as extraction and classification over a domain grounded schema. Stage 1 extracts distinct points from threads; Stage 2 assigns Barrier, Solution, or Metric types together with Pro/Con roles. Trained on 840 curated Deliberatorium examples, CIVICPARSE attains 88.5% accuracy with strong precision (91.1%) and recall (96.5%), substantially outperforming identical prompt only baselines. Beyond the gains from fine tuning, we contribute a reproducible extractor classifier design, a curated dataset, and an evaluation protocol that together cast structured deliberation as a benchmarkable task for AI assisted civic decision making.