Users' Feedback on COVID-19 Lockdown Documentary: An Emotion Analysis and Topic Modeling Analysis

Xiaochuan Shi, Miaoyutian Jia, Jia Li, Quiyi Chen, Liu Guan, Qian LiuFrontiers in Psychology 2022 · DOI

TL;DR

This study applies Natural Language Processing (NLP) techniques—specifically lexicon-based emotion analysis and Latent Dirichlet Allocation (LDA) topic modeling—to decode public sentiment from YouTube comments on a Wuhan lockdown documentary. By mapping emotional shifts and key themes, the researchers demonstrate how granular social media feedback can be used to refine video recommendation systems and health communication strategies.

What Problem Does This Paper Address?

Current video recommendation systems (like those on YouTube or Netflix) often rely on "implicit" feedback, such as watch time and click-through rates, or simple "explicit" feedback like "likes." However, these metrics fail to capture the nuanced emotional resonance or the specific reasons why a user engaged with a piece of content.

In the context of high-stakes health communication—such as documentaries about the COVID-19 pandemic—understanding the audience's psychological response is critical. The authors aim to bridge the gap between raw engagement data and deep sentiment understanding to help marketers and health officials provide more personalized and effective content recommendations.

Methods

The researchers employed a multi-stage computational pipeline using Python 3.6:

  1. Data Cleaning: Standard preprocessing, including the removal of non-English comments and "stop words" (common words like "the" or "this" that carry little semantic value).
  2. Emotion Analysis: The team used the NRC Emotion Lexicon, an authoritative dictionary that maps words to eight basic emotions (anger, fear, anticipation, trust, surprise, sadness, joy, and disgust) and two sentiments (positive and negative). This allowed them to quantify the emotional "pulse" of the comment section.
  3. Topic Modeling (LDA): They applied Latent Dirichlet Allocation (LDA), a generative statistical model that discovers abstract "topics" within a collection of documents. By treating each comment as a mixture of topics and each topic as a mixture of keywords, they identified the primary themes of the conversation.
  4. Visualization: The LDAvis tool was used to project these topics into a two-dimensional space (using multidimensional scaling), allowing researchers to see how distinct or overlapping the different discussion themes were.

Datasets

The study focused on a single, highly influential dataset:

  • Source: YouTube comments from the CGTN documentary "The lockdown: One month in Wuhan."
  • Scale: 1,560 of the "most popular" comments were initially collected.
  • Refinement: After filtering for English-only text, the final dataset consisted of 1,274 comments.
  • Context: At the time of the study, the documentary had over 17 million views and 180,000 likes, making it a significant representative sample of global digital discourse on the Wuhan lockdown.

Key Insights and Ideas

  • Surprising Positivity: Despite the documentary's grim subject matter, the dominant emotions were overwhelmingly positive: Trust (22.8%), Anticipation (17.6%), and Joy (15.4%). This suggests the film successfully pivoted from a tragedy to a narrative of resilience and hope.
  • The "Hero" Narrative: A significant portion of the discourse (over 22%) focused specifically on medical workers as "heroes." This indicates that human-centric storytelling is a primary driver of positive engagement in health documentaries.
  • Sentiment Decay and Fatigue: The researchers noted a temporal shift. While early responses (April 2020) were highly engaged and positive, negative emotions like anger and disgust began to grow faster after July 2020. This likely reflects "pandemic fatigue" and the increasing politicization of lockdown measures globally.
  • Marketing Utility: The authors argue that "trust" and "anticipation" are better predictors for future content recommendations than a simple "like" button. If a user reacts with "trust" to a medical story, the system should prioritize similar human-interest health content rather than just any video tagged "COVID-19."

Key Results

  • Top Themes: Three major themes accounted for 70% of all comments:
  • Stories of fighting the virus (30.1%).
  • Medical workers as heroes (22.8%).
  • Respecting the lockdown/praying for safety (14.2%–19.8%).
  • Temporal Peaks: Engagement and emotional intensity peaked twice: first in April 2020 (during the initial global spread) and again in September 2020.
  • LDA Performance: The model successfully extracted 10 distinct topics which were then categorized into six overarching themes, providing a readable explanation of user opinions that raw data cannot provide.

Limitations and Future Work

  • Text-Only Analysis: The study ignored non-textual feedback such as emojis, images, and video replies, which are increasingly common on social platforms.
  • Language Bias: By filtering for English-only comments, the study likely missed the specific cultural nuances of the local Chinese audience or other non-English speaking regions.
  • Platform Constraints: The researchers noted that YouTube provides limited "explicit" feedback (like nuanced ratings), forcing a reliance on comments which may not represent the "silent majority" of viewers.
  • Future Directions: The authors suggest integrating these NLP models directly into real-time recommendation engines and expanding the analysis to larger, multi-platform datasets to test the scalability of the sentiment-based marketing approach.

Summary updated

Abstract

Conducting emotion analysis and generating users' feedback from social media platforms may help understand their emotional responses to video products, such as a documentary on the lockdown of Wuhan during COVID-19. The results of emotion analysis could be used to make further user recommendations for marketing purposes. In our study, we try to understand how users respond to a documentary through YouTube comments. We chose "The lockdown: One month in Wuhan" YouTube documentary, and applied emotion analysis as well as a machine learning approach to the comments. We first cleaned the data and then introduced an emotion analysis based on the statistical characteristics and lexicon combination. After that, we applied the Latent Dirichlet Allocation (LDA) topic modeling approach to further generate main topics with keywords from the comments and visualized the distribution by visualizing the topics. The result shows trust (22.8%), joy (15.4%), and anticipation (17.6%) are the most prominent emotions dominating the comments. The major three themes, which account for 70% of all comments, are discussing stories about fighting against the virus, medical workers being heroes, and medical workers being respected. Further discussion has been conducted on the changing of different sentiments over time for the ongoing health crisis. This study proves that emotion analysis and LDA topic modeling could be used to generate explanations of users' opinions and feelings about video products, which could support user recommendations in marketing.
BibTeX
@inproceedings{shi2022users,
  author = {Xiaochuan Shi and Miaoyutian Jia and Jia Li and Quiyi Chen and Liu Guan and Qian Liu},
  title = {{Users' Feedback on COVID-19 Lockdown Documentary: An Emotion Analysis and Topic Modeling Analysis}},
  booktitle = {Frontiers in Psychology},
  year = {2022},
  doi = {10.3389/fpsyg.2022.944049},
}
Figure