Analysis and classification of privacy-sensitive content in social media posts

Area of Science:

• Computer Science
• Social Media Analysis
• Natural Language Processing

Background:

• User-generated content often contains private information, posing privacy risks.
• Existing methods for detecting sensitive content rely on unrealistic assumptions about user behavior and platform settings.
• Automatic detection of sensitive information in public online content remains an open challenge.

Purpose of the Study:

• To address the challenge of content sensitivity analysis in user-generated online content.
• To develop and evaluate methods for accurately classifying social media posts as sensitive or non-sensitive.
• To provide a benchmark dataset for future research in content sensitivity analysis.

Main Methods:

• Creation and characterization of a new annotated corpus of approximately ten thousand social media posts, labeled as sensitive or non-sensitive by expert annotators.
• Comparison of content sensitivity analysis with the related task of self-disclosure analysis.
• Implementation and evaluation of several deep neural network models for content sensitivity classification.

Main Results:

• The developed deep neural network models significantly outperform previous naive classification approaches.
• State-of-the-art methods relying on anonymity and lexical analysis demonstrate poor performance in realistic scenarios.
• The new annotated corpus provides valuable data for understanding and addressing content sensitivity.

Conclusions:

• Deep neural networks offer a promising approach for accurate content sensitivity analysis on social media.
• Realistic privacy risk assessment requires methods that do not depend on user anonymity or platform settings.
• Further research using the provided dataset can advance the field of online privacy protection.