TL;DR
This paper tackles the challenge of accurately detecting aggressive online content by moving beyond one-size-fits-all models. They introduce personalized approaches that leverage individual user annotations and conformity measures, demonstrating significant improvements, especially for highly controversial content.
What Problem Does This Paper Address?
Online platforms are rife with content that can be perceived as aggressive, toxic, or offensive. Traditional methods for detecting such content rely on generalized classifiers that treat all users the same. This approach leads to high misclassification rates, particularly for nuanced or controversial texts where individual interpretation plays a significant role. The problem is that a single "correct" label for aggressive content often doesn't exist, as what one person finds offensive, another might not. This paper aims to bridge this gap by developing personalized aggressiveness detection systems that account for individual user beliefs and preferences.
Methods
The authors propose three novel personalization methods:
- Conformity-based Personalization: This method uses simple measures of user conformity, specifically General Conformity (GConf) and Weighted Conformity (WConf). These measures quantify how often a user's annotations align with the majority of other annotators. These conformity scores are then used as input features for a classifier.
- Class-based Embeddings: This approach creates two embeddings for a user based on their annotations of a small set of controversial texts. One embedding represents their annotations of non-aggressive texts, and the other represents their annotations of aggressive texts.
- Annotation-based Embeddings: This method considers individual annotations of a small set of controversial texts. For each annotated text, it uses features like the text's content embedding, its controversy score, and information about the user's rating relative to the majority.
These personalized features are then fed into a classifier (specifically, logistic regression models) to predict aggressiveness.
Datasets
The primary dataset used in this research is the Wikipedia Talk Labels: Aggression dataset. This dataset is notable because it provides individual annotations from multiple crowdworkers for each comment, rather than just a majority vote. This granular information is crucial for developing personalized models. The dataset was split into training, development (dev), and testing sets. Annotators with fewer than a certain number of annotations in the dev/test sets were removed to ensure sufficient data for personalization.
Key Insights and Ideas
- Individual Beliefs Matter: The core idea is that personal beliefs, expressed through annotations, are key to improving aggressiveness detection, especially for subjective content.
- Controversy as a Feature: Highly controversial documents, where annotator disagreement is high, are particularly valuable for extracting user beliefs and driving personalization gains.
- Few Annotations Can Be Enough: Surprisingly, even a small number of annotated controversial documents (as few as 4) can significantly boost personalization performance.
- Personalization Outperforms Generalization: Personalized models consistently outperform generalized models, especially on more controversial texts.
- Trade-off Between Performance and Annoyance: While more annotations generally lead to better performance, there's a point of diminishing returns, suggesting a need to balance performance gains with user experience.
Key Results
- Personalized methods significantly outperform the generalized baseline, with gains being greater for more controversial documents.
- The annotation-based embedding method achieved the best performance, with an F1 score improvement of over 10 percentage points compared to the baseline. For the most controversial 10% of texts, this improvement reached 22.7 percentage points.
- The conformity-based method also showed substantial gains, particularly for the most controversial texts, where the non-personalized method was "completely lost."
- Even with just 4 previously annotated controversial texts, personalized models achieved performance gains of up to 7.3 percentage points over the baseline.
- The training time for personalized models, while longer than the baseline, was manageable, with annotation-based methods being the most computationally intensive.
Limitations and Future Work
- Dataset Limitations: The authors highlight the need for new datasets that provide individual human annotations rather than aggregated labels, to better capture diverse beliefs.
- Computational Cost: While the models are efficient, the annotation-based method can become computationally expensive with a large number of input features.
- User Annoyance: The paper acknowledges the need to balance the number of required annotations with user annoyance.
- Future Work: The authors suggest applying these personalization methods to other NLP tasks involving subjective content (e.g., hate speech, cyberbullying, emotion detection, sentiment analysis) and exploring more advanced language representations like XLNet and RoBERTa. They also plan to address the issue of imbalanced classes in aggressiveness detection.
Summary updated