Back to top

Sentiment Analysis

  • Introduction
  • Exploratory data analysis
  • Data Visualization
  • Methodology
  • Sentiment Analysis
  • Conclusion
  • Code

    • Sentiment Analysis

    Sentiment Score

    The Sentiment Score of other rumor content is shown in Figure 1. The analyzed content exhibits an average sentiment score of 0.34, approaching the neutral threshold but categorized as “negative” within the applied classification framework. Score distribution follows a near-normal pattern, with a peak density around 0.34. Notably, 62.1% of scores fall within the 0.3–0.4 range (classified as moderately negative), while 23.7% cluster in the 0.2–0.3 interval (strongly negative). Only 14.2% exceed 0.45, indicating limited positive polarity.

    This skewed distribution toward the lower spectrum aligns with the overall “negative” classification. To operationalize these findings, we recommend:

    • Identifying dates with sentiment minima (e.g., scores ≤0.2) through temporal granularity analysis.
    • Conducting content audits on these critical periods to isolate linguistic or thematic triggers of extreme negativity.
    • Developing mitigation strategies such as targeted counter-messaging or platform-level sentiment-dampening interventions.
    Figure 1 Sentiment Score Distribution of Rumor Post.


    The sentiment distribution of non-rumor posts in Figure 2 reveals a concentration of scores around 0.36, classified as “negative” by the applied model. This classification raises a critical paradox: despite the posts’ fact-checking purpose (debunking misinformation), their aggregated sentiment skews negatively. A plausible explanation lies in their content composition—the frequent use of inherently negative terms during refutation, coupled with discussions of socially concerning topics.


    Figure 2 Sentiment Score Distribution of Non-Rumor Posts.


    Methodologically, the analysis employed a BERT-based model pre-trained on generic corpora. While BERT outperforms lexicon-based tools like SnowNLP in semantic precision, its limited exposure to Weibo-specific linguistic patterns (e.g., colloquial negations, sarcastic refutations) may lead to misalignment. For instance, neutral/educational phrases contextualized within debates about misinformation might inherit unintended negative polarity due to the model’s inability to fully adapt to platform-specific pragmatics.

    Daily Sentiment Score

    Analysis of the temporal sentiment distribution (left figure in Figure 3) identifies 2022-01-30 as the date with the lowest daily average sentiment score (0.34), indicating peak negativity.

    Figure 3 Daily Sentiment Score and Content of Rumor Post.

    A targeted examination of this day’s user comments reveals heightened expressions of anxiety and distrust, exemplified by recurring phrases such as “uncertain risks” and “lack of official clarification.”

    This concentration of negative sentiment likely stems from two factors:

    1. Information vacuum: The Proliferation of unverified claims during critical events amplified public uncertainty.
    2. Emotional contagion: Anxiety-laden comments reinforced collective panic through social reinforcement.

    To address such scenarios, we propose prioritized interventions on dates with sentiment minima:

    • Timely refutation: Release authoritative clarifications matching rumor themes.
    • Preemptive education: Disseminate science-based narratives about rumor identification before crisis triggers.