Neural Network Prediction of Censorable Language

Kei Yin Ng, Anna Feldman, Jing Peng, Chris LeberknightNAACL 2019

TL;DR

This paper tackles the challenge of automatically detecting censorable language on social media, specifically Sina Weibo, by building a neural network classifier. They demonstrate that linguistic features alone can predict censorship with high accuracy, suggesting these features could be leveraged for censorship circumvention.

What Problem Does This Paper Address?

The free flow of information is crucial for democratic societies, yet political censorship is a widespread phenomenon. Platforms like Sina Weibo, China's dominant microblogging service, employ sophisticated censorship systems that restrict content. While previous research has focused on technical methods for censorship evasion or identifying what gets censored, this paper addresses the gap of automatically predicting if a post is likely to be censored based on its linguistic characteristics. Understanding these linguistic "fingerprints" is key to both improving censorship detection and potentially developing more subtle circumvention strategies.

Methods

The core of this paper is the development of a neural network classifier to predict censorable language. They extract a variety of linguistic features from social media posts, including:

  • LIWC features: These analyze text based on predefined linguistic categories (e.g., pronouns, social processes, cognitive processes) to capture psychological and stylistic aspects.
  • CRIE features: This tool analyzes Chinese texts for readability, syntax, semantics, and cohesion.
  • Sentiment features: Using BaiduAI, they obtain positive and negative sentiment scores for each post.
  • Semantic features: Leveraging a Chinese thesaurus, they extract semantic classes and derive a semantic ambiguity feature.
  • Frequency & Readability features: They compute character and word frequencies and combine them with word count to assess text readability.
  • Follower count: The number of followers of the post's author is also used as a feature.

They then train a Multilayer Perceptron (MLP) classifier using these features. Notably, they explore different combinations of features, finding that a mix of CRIE, sentiment, semantic, frequency, readability, and follower features (excluding LIWC) yields the best performance.

Datasets

The authors utilize two primary datasets:

  1. Newly scraped Sina Weibo data: This dataset was collected by tracking posts that mention "sensitive" topics or are authored by "sensitive" users. They developed a novel web scraping and decoding method to identify censored posts, distinguishing between system-deleted and user-deleted content. This process resulted in approximately 41,000 posts, with 952 (2.28%) identified as censored.
  2. Pre-existing corpus from Zhu et al. (2013): This dataset consists of approximately 20,000 text-only posts from around 3,500 sensitive users, with 930 (4.63%) identified as censored.

Key Insights and Ideas

  • Linguistic fingerprints of censorship: The paper strongly suggests that censorable language leaves discernible linguistic traces. This is a significant finding, moving beyond topic-based detection to stylistic and semantic analysis.
  • Predictive power of linguistic features: The neural network classifier achieved high accuracy using only linguistic features, indicating that the "how" of writing is as important as the "what" in determining censorship.
  • Potential for censorship circumvention: The authors hypothesize that understanding these linguistic features could be used to develop more sophisticated censorship circumvention techniques, potentially by subtly altering writing style to avoid detection.
  • Importance of specific feature types: The study highlights that features related to writing style, readability, sentiment, and semantic complexity are particularly crucial for distinguishing censored from uncensored posts, more so than broad vocabulary categories captured by LIWC alone.

Key Results

The neural network classifier achieved an accuracy of 88.50% using a combination of linguistic features (specifically, CRIE, sentiment, semantic, frequency, readability, and follower features). This performance is significantly higher than previous human baselines, demonstrating the effectiveness of their automated approach. They also found that using all features except LIWC yielded the best results, suggesting that the broad, general vocabulary categories in LIWC might not be as discriminative as more nuanced linguistic features.

Limitations and Future Work

The authors acknowledge that their work is based on data collected from Sina Weibo, and the findings might not directly generalize to other platforms or languages without further investigation. They also note that while their model performs well, it does not achieve perfect accuracy.

Future work could explore:

  • Applying these linguistic feature extraction and classification methods to other social media platforms and in different linguistic contexts.
  • Investigating how these identified linguistic features can be actively used for censorship circumvention.
  • Further refining the feature extraction process to capture even more subtle linguistic nuances.

Summary updated

BibTeX
@inproceedings{ng2019neural,
  author = {Kei Yin Ng and Anna Feldman and Jing Peng and Chris Leberknight},
  title = {{Neural Network Prediction of Censorable Language}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2019},
}
Figure