The effect of information controls on developers in China: An analysis of censorship in Chinese open source projects

Jeffrey Knockel, Masashi Crete-Nishihata, Lotus RuanCOLING 2018

TL;DR

This paper investigates how Chinese censorship influences individual developers by analyzing keyword blacklists within open-source projects hosted on GitHub. They discovered over 1,000 such blacklists, revealing a surprisingly low overlap between them, which challenges assumptions about centralized censorship and suggests independent curation or diverse sources.

What Problem Does This Paper Address?

While the impact of China's intermediary liability laws on large commercial platforms is well-documented, the effect of these censorship pressures on individual developers remains largely unexplored. This paper aims to bridge that gap by examining whether individual developers incorporate censorship features, like keyword filtering, into their open-source projects. Understanding this could shed light on the motivations behind such inclusions and the methods used to create these censorship lists, particularly in the context of China's complex and often decentralized regulatory environment.

Methods

The researchers developed a novel heuristic-based tool to extract lists of strings from various structured data formats commonly found in GitHub projects. This tool was designed to be robust and avoid overfitting. They then employed a two-pronged approach to identify Chinese blacklists:

  1. Naive Approach: This involved searching for lists containing the specific keyword "法轮" (Falun), a known sensitive term in China. Manually verifying these lists and then inspecting their associated files for other potential blacklists helped expand the dataset.
  2. Machine Learning Approach: To overcome the limitations of the naive approach, they used a one-class Support Vector Machine (SVM). This classifier was trained on known Chinese blacklists and then used to identify potential blacklists from the extracted lists. The dimensionality of the data was reduced using singular-value decomposition.

Datasets

The primary dataset for this study consists of keyword blacklists extracted from open-source projects hosted on GitHub. The researchers collected over 1,000 unique Chinese blacklists, containing more than 200,000 unique keywords. This represents the largest dataset of Chinese blacklisted keywords to date. The training data for the machine learning model was derived from publicly available Chinese blacklists and a specific list discovered in Google's Javascript code for its mainland Chinese search engine.

Key Insights and Ideas

  • Decentralized Censorship: The significant lack of keyword overlap between the discovered blacklists strongly suggests that censorship is not a monolithic, top-down system. Instead, it points towards independent curation by individual developers or diverse, uncoordinated sources.
  • Developer Motivations: The presence of these blacklists implies that developers either feel pressure (formal or informal) to implement censorship or are accustomed to such requirements for market access.
  • Scale of Curation: The sheer size of many blacklists (over 1,000 keywords) makes individual manual curation seem unlikely, raising questions about how these extensive lists are generated and shared.
  • Beyond Commercial Platforms: This research extends the understanding of censorship beyond large commercial entities to the individual developer level within the open-source ecosystem.

Key Results

  • The study successfully scraped GitHub and extracted 45,986 lists containing at least 20 strings and at least one Chinese character.
  • The naive approach identified 915 potential Chinese blacklists, which were reduced to 838 true blacklists after manual verification. An additional 46 were found through manual inspection of associated files, totaling 884.
  • The machine learning approach positively classified 1,391 lists, of which 1,054 were verified as true Chinese blacklists.
  • Collectively, the identified blacklists comprised over 215,007 unique keywords.
  • The distribution of keywords per list varied, with a mean length of approximately 2,128 keywords and a median of 1,026.
  • Common keywords included terms related to prurient interests, Falun Gong, political movements (e.g., Tibetan independence), government criticism, and political leaders.

Limitations and Future Work

  • Limitations: The paper acknowledges that the naive approach is limited to detecting blacklists containing "法轮" or those in files with such lists. The machine learning approach, while more sophisticated, still required manual verification of all positive classifications, which can be time-consuming. The size of the training set for the machine learning model was limited by publicly available data.
  • Future Work: The authors suggest an interview study with developers to gain deeper insights into their motivations for including these lists and their curation processes. Further research could explore how these lists are shared and updated over time by analyzing GitHub commit histories. Additionally, sampling all China-based projects, regardless of whether they perform censorship, could provide a broader characterization of filtering practices. Analyzing GitHub's social networking data (forks, stars, watchers) might also reveal how these blacklists are shared.

Summary updated

BibTeX
@inproceedings{knockel2018effect,
  author = {Jeffrey Knockel and Masashi Crete-Nishihata and Lotus Ruan},
  title = {{The effect of information controls on developers in China: An analysis of censorship in Chinese open source projects}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2018},
}
Figure