Characterizing News Portrayal of Civil Unrest in Hong Kong, 1998–2020

James Scharf, Arya D. McCarthy, Giovanna Maria Dora DoreACL 2021

TL;DR

This paper uses natural language processing to analyze how Western and Hong Kong-based news outlets portrayed civil unrest in Hong Kong from 1998-2020. They found significant differences in topic coverage and word usage, suggesting that media framing can vary substantially based on the source's origin, particularly concerning topics like democracy and police violence.

What Problem Does This Paper Address?

Understanding how civil unrest is portrayed in the media is crucial for grasping public perception and the dynamics of social movements. However, existing research often focuses on single events or limited timeframes. This paper addresses the need for a longitudinal analysis of media coverage of civil unrest, specifically in Hong Kong, a region with a complex political history. The authors aim to uncover differences in how Western and Hong Kong-based English-language news sources frame these events, exploring whether these differences are consistent over time or shift with major political developments. This is important because media framing can influence public opinion, legitimize or delegitimize protests, and shape international understanding of events.

Methods

The researchers employed a multi-faceted approach using natural language processing (NLP) techniques:

  • Lexical Frequency Analysis (ANOVA): They analyzed the frequency of 19 selected keywords related to protests (e.g., "democracy," "protest," "confrontation") across different news sources. ANOVA was used to statistically determine if differences in keyword frequency were significant.
  • Topic Modeling (LDA): Latent Dirichlet Allocation (LDA) was used to identify and characterize the main topics discussed in the news articles. This allowed them to see which themes were prevalent in different sources and over time.
  • Lexical Usage Analysis (Word Embeddings): To understand how words are used, not just if they are used, they employed word embedding models. By comparing the "neighborhoods" (contextual words) of keywords in Western vs. Hong Kong-based corpora, they could detect shifts in word meaning and usage over time and across different media origins.
  • Sentiment Analysis (BERT-based): A BERT-based model was used to classify sentences as positive or negative, providing an overall sentiment score for each article. This helps to understand the general tone of coverage.

Datasets

The study utilized a corpus of 4175 English-language newspaper articles collected from six Western-based sources (The New York Times, The Wall Street Journal, The Washington Post, The Financial Times, The Guardian, The Times) and two Hong Kong-based sources (The China Daily, The South China Morning Post). Articles were collected using keyword searches related to "Hong Kong" and "protests," "rallies," "marches," or "riots," and had to be at least 300 words long. The data spans from January 1998 to June 2020.

Key Insights and Ideas

  • Divergent Framing: The most striking insight is the consistent and significant difference in how Western and Hong Kong-based media portray civil unrest. This isn't just about what's covered, but how it's framed.
  • "Protest" and "Democracy" Under-discussed Locally: Hong Kong-based sources discussed "protest" and "democracy" less frequently than Western sources, despite covering the events. This could indicate a form of self-censorship or a strategic choice to appeal to a local audience.
  • Police Violence: A Western Focus: Western media significantly increased its reporting on "police violence" around the 2019 Anti-Extradition Law Amendment Bill Movement, while Hong Kong-based media showed a much more muted response to this specific topic.
  • Shifting Meanings of "Confrontation": The meaning of words like "resistance," "confront," and "confrontation" evolved significantly, especially after July 2019. Western media increasingly associated "resistance" with protesters' emotions, while Hong Kong media linked it to adversarial behavior, highlighting a dichotomous framing.
  • Polarizing Language: The study found evidence of semantic divergence in the usage of polarizing terms like "tension" and "severe," suggesting different approaches to framing protest events.

Key Results

  • Lexical Frequency: ANOVA revealed statistically significant differences in the frequency of 15 out of 19 selected keywords between Western and Hong Kong-based articles. Keywords like "democracy," "protest," and "protests" showed the largest disparities.
  • Topic Representation: While both source types covered similar events, the prominence of topics differed. For instance, "Marches/Rallies" and "Bills" were more prevalent in Hong Kong-based media, while Western media focused more on "Police Violence."
  • Lexical Usage: The analysis of word embeddings showed notable semantic differentiation. For example, "resistance" had a low similarity score, indicating a significant shift in its usage. The study identified July 2019 as a turning point for the meaning of several keywords.
  • Sentiment: Sentiment analysis showed a generally pessimistic tone across all sources, with Hong Kong-based sources exhibiting slightly more negative sentiment on average.

Limitations and Future Work

  • Sample Selection: The authors acknowledge the importance of sample selection and how it can influence findings. They also note that their analysis of "protest" might be affected by its lower prevalence in Hong Kong-based sources, potentially due to self-censorship.
  • Causality: The study identifies correlations and differences but does not establish causality. Future work could explore whether news language causes protest sizes or if there's a Granger-causal relationship.
  • Temporal Granularity: The analysis often bins articles by year, which can obscure finer temporal shifts. Future work could focus on more precise measurement of news content and framing around specific protest events.
  • Sentiment Nuance: The sentiment analysis provides a binary positive/negative classification, which obscures the reason for the sentiment (e.g., opposing protests vs. opposing police response).
  • Event Extraction: The authors are working to quantify how framing differences might "stymie" event extraction systems, a crucial practical application of their research.

Summary updated

BibTeX
@inproceedings{scharf2021characterizing,
  author = {James Scharf and Arya D. McCarthy and Giovanna Maria Dora Dore},
  title = {{Characterizing News Portrayal of Civil Unrest in Hong Kong, 1998–2020}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure