Proposed architecture for product review summarisation: a modular pipeline for aspect-based insight extraction and synthesis

Вантажиться...
Ескіз

Дата

2025

Науковий керівник

Назва журналу

Номер ISSN

Назва тому

Видавець

КПІ ім. Ігоря Сікорського

Анотація

In the domain of e-commerce, user reviews are a critical resource for decision-making. However, the sheer volume and unstructured nature of these reviews present a significant challenge for consumers seeking actionable insights. Traditional extractive summarization methods often fail to capture nuance, while end-to-end Large Language Model (LLM) approaches struggle with hallucinations and lack of structural control. This paper proposes a novel, linear, multi-stage architecture that transforms unstructured text into the Quantified Aspect-Based Summary (QABS). The proposed pipeline utilizes a modular approach, integrating Coreference Resolution, Low-Rank Adaptation (LoRA) finetuned models for tuple extraction, and dynamic topic modeling. By decomposing reviews into atomic insights and re-synthesizing them using blueprint prompting, this architecture ensures high clarity, trust, verifiability, and quantification of user sentiment.

Опис

Ключові слова

Aspect-Based Sentiment Analysis (ABSA), Large Language Models, LoRA, Coreference Resolution, Dynamic Topic Modeling, Product Summarization

Бібліографічний опис

Panasiuk, Ya. Proposed architecture for product review summarisation: a modular pipeline for aspect-based insight extraction and synthesis / Yaroslav Panasiuk // Системні науки та інформатика : збірка доповідей ІV науково-практичної конференції, [Київ], 1–5 грудня 2025 р. / Навчально-науковий Інститут прикладного системного аналізу КПІ ім. Ігоря Сікорського. – Київ, 2025. – С. 288-291.

ORCID

DOI