TL;DR
This paper introduces two novel, large datasets of moderated news comments from Slovenia and Croatia, addressing the scarcity of resources for non-English abusive language detection. They analyze comment deletion patterns across users, threads, and positions, finding that deletion is less random in Slovenia and that specific users are more frequently targeted.
What Problem Does This Paper Address?
The proliferation of online user-generated content has led to an explosion of potentially inappropriate comments, ranging from hate speech to spam. Effectively moderating this content at scale is a significant challenge. While machine learning approaches show promise, they heavily rely on large, well-annotated datasets. However, existing datasets are predominantly in English, leaving a critical gap for researchers and practitioners working with other languages and cultural contexts. This paper aims to bridge this gap by creating and analyzing datasets for Slovene and Croatian, two languages underrepresented in this research area.
Methods
The authors introduce two new datasets: one from the Slovene news portal MMC RTV and another from the Croatian news portal 24sata. These datasets contain not only the comments themselves but also crucial metadata, most importantly, whether a comment was deleted by moderators. To enable research while protecting privacy and sensitive content, the datasets are published in a pseudo-anonymized and encrypted format. They also provide baseline fastText classification models trained on the non-encrypted data.
The paper's analysis involves:
- Manual Annotation: A subset of deleted comments was manually classified into categories of inappropriate content (e.g., discrimination, insult, swearing) based on a pre-existing schema.
- Statistical Analysis: They performed statistical analyses to understand deletion patterns, comparing observed distributions of deleted content across users, threads, and comment positions against random distributions. This involved using the Kolmogorov-Smirnov test and Wasserstein distance to quantify deviations from randomness.
- Baseline Model Training: FastText classifiers were trained on the datasets to identify potentially inappropriate content, with hyperparameter tuning to optimize performance.
Datasets
- Slovenian MMC RTV Dataset: Comments from the MMC RTV web portal, spanning from 2010 to 2017.
- Croatian STY (24sata) Dataset: Comments from the news portal 24sata, covering 2007 to 2017.
Both datasets are substantial, containing millions of comments and hundreds of millions of words. They are made available in encrypted form, with pseudo-anonymized metadata, and baseline fastText models are also provided.
Key Insights and Ideas
- Cross-Lingual Moderation Differences: The analysis reveals significant differences in moderation policies and the nature of deleted content between Slovenia and Croatia. Croatian deleted comments tend to be more overtly aggressive (insults, swearing), while Slovene comments are often more subtle (irony, off-topic remarks, targeting moderators).
- Non-Random Deletion Patterns: Deletion of comments is not a purely random process. The study demonstrates that specific users, threads, and even comment positions exhibit statistically significant deviations from random deletion patterns.
- User-Centric Deletion: The most pronounced non-randomness is observed at the user level, suggesting that certain users are disproportionately targeted for comment deletion. This supports the idea that "inappropriate comments are not a blunder that happens to everyone now and then but that there are consistently 'non-conforming' users."
- Thread and Location Influence: While less pronounced than user-level patterns, thread-level deletion also deviates from randomness. Interestingly, comments at the beginning of a thread appear slightly more prone to deletion than those in the middle or at the end, contrary to the expectation that discussions might escalate and become more heated later.
- Dataset Value: The creation and release of these datasets are crucial for advancing research in abusive language detection for under-resourced languages, enabling more inclusive and culturally aware NLP applications.
Key Results
- Dataset Size: The combined datasets contain nearly 25 million comments and over 700 million words.
- Deletion Rates: The MMC (Slovenian) portal deleted over 8% of texts, while STY (Croatian) deleted under 2%. However, nearly half of MMC users had at least one comment deleted, compared to under 10% for STY.
- Statistical Significance: Kolmogorov-Smirnov tests consistently showed statistically significant differences between observed and random deletion distributions across users, threads, and locations for both datasets.
- Model Performance: Baseline fastText models achieved a ROC AUC score of approximately 0.794 for the MMC dataset and 0.793 for the STY dataset, indicating moderate predictive power. Adding character n-grams significantly improved performance.
Limitations and Future Work
- Encryption: The datasets are encrypted, limiting direct application of models to real-world scenarios without decryption. The authors provide baseline models trained on non-encrypted data as a workaround.
- Annotation Schema: A notable portion of deleted comments (19% in MMC, 18% in STY) did not fit the predefined annotation categories, suggesting the schema might need refinement for broader applicability.
- Model Performance: The authors acknowledge that the achieved ROC AUC scores are "far from satisfying" and leave the task of building stronger prediction models to future work.
- Future Directions: The paper suggests further research into building more robust prediction models, exploring the cultural and publisher-specific reasons behind the observed deletion patterns, and potentially refining annotation schemas.
Summary updated