Pattern-Guided Implicit Sentiment Analysis: Knowledge-Augmented Prompting with Aspect-Sentiment Joint Extraction for Automotive User Feedback

2026-99-0753

5/15/2026

Authors
Abstract
Content
Implicit sentiment analysis of automotive user feedback is crucial for understanding user opinions. Automotive user feedback often express opinions in an indirect way and are accompanied by a dense array of industry terms. Therefore, without costly fine-tuning, both aspect identification and sentiment analysis are rather difficult. We propose a Pattern-Guided pipeline for implicit sentiment analysis to achieve the joint extraction of aspect and sentiment. This pipeline first performs Pattern Anchoring, mapping colloquial expressions and slang to the standardized vehicle component knowledge system. Then, using Knowledge-Augmented Prompting, these domain rules are injected into well-designed prompt templates. In this pipeline, the large language model (LLM) is applied to output JSON records suitable for comprehending, including aspects, sentiments, confidence levels, and brief reasons. To enhance stability, we employ an improved prompt and consistency-driven confidence fusion to generate multiple JSON records with confidence-building rules. The generated JSON records further enter the Attribution & Aggregation layer, which is used to cluster negative feedback, merge synonymous expressions, and generate traceable query summaries and priority signals. On a corpus containing 2,062 Chinese samples, the 7B open-source model qwen1.5-7b-chat achieved an accuracy rate of 94.7%, outperforming supervised fine-tuning at 93.1% and direct classification at 92.7%, and approaching qwen-plus at 95.7%. In a case study focusing on the Volkswagen Magotan, this method compressed 137 negative comments into traceable and sortable problem clusters. The code and data are available at https://github.com/ppg94/Implicit-Sentiment-Joint-Extraction-Automotive.
Meta TagsDetails
DOI
https://doi.org/10.4271/2026-99-0753
Citation
Chang, G., Deng, Z., Ma, A., Yao, J., et al., "Pattern-Guided Implicit Sentiment Analysis: Knowledge-Augmented Prompting with Aspect-Sentiment Joint Extraction for Automotive User Feedback," Interntional Conference on the New Energy and Intelligent Vehicles, Hefei, China, November 2, 2025, https://doi.org/10.4271/2026-99-0753.
Additional Details
Publisher
Published
May 15
Product Code
2026-99-0753
Content Type
Technical Paper
Language
English