Abstract:
"Analysing online product reviews for both aspect sentiment and relationships between entities
presents a challenge. Existing models often excel in one area but struggle in the other, limiting
their comprehensive understanding of user opinions. This limitation can hinder accurate
sentiment analysis and potentially mislead users.
This work proposes MUTA (Mistral-T5 Unified Analysis): a novel hybrid large language model
(LLM) approach that addresses this challenge. MUTA leverages the strengths of two models:
1. T5 small (fine-tuned): Summarizes product reviews, capturing the essence of user
opinions.
2. Mistral-7b: Analyzes the summarized text to identify crucial components, extract
relationships between them (particularly ""Customer-Manufacturer"") and assign severity
scores.
MUTA's effectiveness is demonstrated through a multi-faceted evaluation. T5 summaries
provide concise overviews of user opinions in reviews, leading to improved understanding of
overall sentiment. Mistral, utilizing these summaries, excels at extracting vehicle components
and uncovering relationships between them, particularly focusing on the crucial ""CustomerManufacturer"" dynamic. Additionally, MUTA achieves a high Rogue score, a metric specifically
designed for evaluating summarization tasks"
