Scholarlink

<p>The rapid development of autonomous driving technology has raised widespread concerns about its safety and reliability. This study proposes a method integrating deep learning and domain knowledge graphs to systematically identify user sentiment and infer safety requirements from 52,847 user comments collected from China’s four major social media platforms. We constructed an autonomous driving safety knowledge graph containing 1,624 entity nodes and 5,873 relationship edges, integrating ISO 26262 and SAE J3016 standards. Using BERTopic for topic modeling and RoBERTa for sentiment analysis, we achieved an overall classification accuracy of 95.3% and macro-average F1-score of 95.3% and identified 8 major discussion topics and 417 unique safety requirement entities. Results show that 43.8% of comments expressed negative sentiment, significantly higher than general electric vehicle research (22.6%). Sensor reliability (67.8% negative), emergency braking systems (61.2% negative), and system transparency (58.9% negative) emerged as the three major safety concerns. The top three high-frequency requirements were improving sensor accuracy in adverse weather, enhancing obstacle recognition, and improving nighttime perception, with 3,223 cumulative mentions. Cross-platform analysis revealed that professional automotive forums showed significantly higher negative sentiment (47.2%, 45.8%) than social media platforms, reflecting differences in user expertise. Ablation studies confirm that the knowledge graph module contributes measurable gains in requirement completeness and failure mode discovery compared to a pipeline without graph-based reasoning. This study provides methodological contributions for autonomous driving acceptance research and offers data-driven decision support for manufacturers, policymakers, and researchers.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid deep learning and knowledge graph framework for automated safety requirement inference: Combining BERTopic, RoBERTa, and graph-based reasoning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月30日

查看全文
  • KunYang Liu,
  • YiMing Sun,
  • YaXiao Chen

摘要

The rapid development of autonomous driving technology has raised widespread concerns about its safety and reliability. This study proposes a method integrating deep learning and domain knowledge graphs to systematically identify user sentiment and infer safety requirements from 52,847 user comments collected from China’s four major social media platforms. We constructed an autonomous driving safety knowledge graph containing 1,624 entity nodes and 5,873 relationship edges, integrating ISO 26262 and SAE J3016 standards. Using BERTopic for topic modeling and RoBERTa for sentiment analysis, we achieved an overall classification accuracy of 95.3% and macro-average F1-score of 95.3% and identified 8 major discussion topics and 417 unique safety requirement entities. Results show that 43.8% of comments expressed negative sentiment, significantly higher than general electric vehicle research (22.6%). Sensor reliability (67.8% negative), emergency braking systems (61.2% negative), and system transparency (58.9% negative) emerged as the three major safety concerns. The top three high-frequency requirements were improving sensor accuracy in adverse weather, enhancing obstacle recognition, and improving nighttime perception, with 3,223 cumulative mentions. Cross-platform analysis revealed that professional automotive forums showed significantly higher negative sentiment (47.2%, 45.8%) than social media platforms, reflecting differences in user expertise. Ablation studies confirm that the knowledge graph module contributes measurable gains in requirement completeness and failure mode discovery compared to a pipeline without graph-based reasoning. This study provides methodological contributions for autonomous driving acceptance research and offers data-driven decision support for manufacturers, policymakers, and researchers.