AMOS: Active Mining with Observable Steps for Agenda-Driven Review Knowledge Discovery

Résumé fourni par la source

Online reviews inform consumer decisions, but the right decision is agenda relative: the same restaurant may pass for value yet fail on allergy safety, since each agenda defines different evidence, thresholds, and rules. Existing methods summarize opinions but neither execute agenda-specific decision logic nor ground verdicts in cited reviews, while end-to-end LLM prompting risks hallucinations. We formulate Agenda-Driven Review Knowledge Discovery (ARKD) and propose Active Mining with Observable Steps (AMOS), a two-phase framework: ASC compiles natural-language agendas into deterministic Agenda Specs, and TEKOA applies our Foveal Search principle, where lightweight retrieval probes scan broadly while costly LLM verification focuses on prioritized candidates with Bayesian-calibrated early stopping. On ARKD-Bench built from Yelp data, AMOS skips 26% of reviews, improves Macro-F1 by up to 47% over the strongest LLM baseline, and yields auditable verdict traces.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.