An LLM-based multi-agent system for multilingual product review analysis

Product reviews influence consumer understanding and purchasing decisions, but their volume, diversity, and multilingual content make manual analysis difficult. Users must often synthesize unstructured feedback concerning sentiment, product features, usability, reliability, delivery, support, and other practical dimensions. Existing review-analysis systems commonly address classification, summarization, or visualization separately, providing limited support for multilingual processing, review-grounded reasoning, structured querying, and visual analytics within one workflow. In this paper, we propose an orchestrated LLM-based multi-agent system for multilingual product review analysis. The six-agent workflow integrates review retrieval and SQL-backed storage, language detection and translation, topic assignment, semantic tagging, retrieval-grounded semantic reasoning, and constrained analytics and visualization. Reviews are cached in product-specific SQLite tables and enriched with language, topic, and semantic fields, enabling natural-language queries and review-grounded textual or visual responses. The system is evaluated using controlled product-review case-study datasets, fixed prompt benchmarks, an OATS-derived task-level benchmark, and a held-out MARC cross-language set, together with cost and latency measurements. In the primary application benchmark, the proposed workflow satisfied all success criteria for 24/36 prompts (66.7%), compared with 14/36 (38.9%) for the conventional pipeline, an absolute improvement of 27.8 percentage points ( p = 0.021 ). It supported a broader range of tested requests, particularly multilingual and paraphrased queries, synthesis across reviews, analytics, and context-dependent follow-up requests. The held-out MARC evaluation yielded exact topic agreement of 0.840, semantic exact-set agreement of 0.684, and mean semantic Jaccard similarity of 0.862, indicating cross-language consistency under the tested conditions. These results support the value of orchestrating structured storage, multilingual enrichment, retrieval-grounded reasoning, and visual analytics within a unified workflow, while the conventional pipeline retained advantages in latency, cost, and selected boundary cases. The configurable architecture also provides a foundation for adaptation to other review-based domains through domain-specific data sources, schemas, taxonomies, and validation targets, with cross-domain evaluation remaining future work.