Large language model-assisted public opinion analysis of emergency events of major transportation infrastructure through social media: A case study of the Humen bridge incident

Major transportation infrastructure emergencies often generate widespread attention and discussion across social media. Failing to respond effectively to public opinion can exacerbate crises and undermine infrastructure trust and reputation. This study examines public opinion on Weibo surrounding the Humen Bridge vortex-induced vibration incident (May 5–18, 2020), employing the lexicon-based approach, SnowNLP, and five BERT models for sentiment analysis, alongside LDA and BERTopic for topic modeling. Results indicate that 82.2% of 21,061 comments were made within 72 h (May 5–7). The weighted lexicon approach showed limited effectiveness (F1 = 0.4466). Domain-adapted SnowNLP (F1 = 0.7195) outperformed RoBERTa-wwm-ext and StructBERT but remained below BERT-base-chinese (F1 = 0.7616). An LLM-assisted annotation experiment showed 74.4%–79.1% agreement with human labels, but BERT trained on these labels achieved F1 = 0.7075 (zero-shot) and F1 = 0.7021 (few-shot), a 7–8% reduction from the human-labeled baseline. LDA identified seven topics led by Vortex-Induced Vibration Discussion (29.1%) and Expert Authority Trust (14.5%). BERTopic identified 19 topics with Expert Skepticism as the largest (14.6%). Overall negativity reached 55.6%, with daily negativity decreasing from 62.7% on May 5 to 54.2% on May 6, before rebounding to 56.7% on May 7. Expert Authority Trust increased from 10.9% on May 5 to 18.5% on May 7 with negativity rising from 54.3% to 64.0%, while Public Shock Response declined from 19.0% to 6.8% with negativity decreasing from 87.1% to 67.0%. These findings provide practical guidance for crisis communication in major transportation infrastructure emergencies.