TL;DR
Cyberbullying detection models often struggle with sarcasm, leading to misclassifications. This paper proposes a method to improve cyberbullying detection by incorporating a sarcasm detection score into a multilayer perceptron (MLP) model, achieving better accuracy and F1 scores than existing BERT-based approaches.
What Problem Does This Paper Address?
The internet has become a primary platform for communication, but this has also led to a rise in cyberbullying, particularly affecting minority groups. While deep learning models like BERT have shown promise in detecting cyberbullying, they often fail to correctly classify texts that employ sarcasm. Sarcasm, where the intended meaning is the opposite of the literal words, can make a text appear innocent when it's hostile, or vice-versa, confusing detection systems. This ambiguity poses a significant challenge for automated cyberbullying detection, leading to false positives and false negatives. The paper aims to bridge this gap by explicitly addressing the impact of sarcasm on cyberbullying detection accuracy.
Methods
The authors first evaluate several existing BERT-based cyberbullying detection models: vanilla BERT, HateBERT, and CyberBERT. Following this, they conduct a failure analysis to identify common misclassification patterns, pinpointing sarcasm as a key issue. To address this, they propose a novel approach that integrates a sarcasm detection layer into their cyberbullying detection pipeline.
Specifically, they use a pre-trained CNN-LSTM based sarcasm detection model (achieving 0.829 accuracy on a separate dataset). This sarcasm score, along with the output from a fine-tuned HateBERT model and the BERT embedding of the input text, are fed into a simple multilayer perceptron (MLP) with two hidden layers. This MLP acts as the final cyberbullying classifier. The MLP is trained using accuracy as the metric and Adagrad as the optimizer. They also perform an ablation study to isolate the impact of the sarcasm detection component versus the MLP itself.
Datasets
The paper utilizes three datasets for cyberbullying detection:
- Twitter: A dataset with 7945 cyberbully and 38072 non-cyberbully instances.
- YouTube: A dataset with 417 cyberbully and 3047 non-cyberbully instances.
- Kaggle: A dataset with 2806 cyberbully and 5993 non-cyberbully instances.
For sarcasm detection, they use a separate dataset gathered by Shmueli et al. (2020), consisting of 15,000 sarcastic and 15,000 non-sarcastic tweets.
Key Insights and Ideas
- Sarcasm is a significant hurdle for cyberbullying detection: The failure analysis clearly shows that sarcastic texts are frequently misclassified by state-of-the-art BERT models, contributing to both false positives and false negatives.
- Integrating sarcasm detection can boost performance: Explicitly incorporating a sarcasm score as a feature into a cyberbullying classifier can lead to substantial improvements in accuracy and F1 scores.
- MLPs are effective for feature combination: A simple MLP can effectively weigh and combine different features (BERT embeddings, HateBERT output, sarcasm score) to make a more robust cyberbullying classification.
- Dataset imbalance matters: The performance of models is significantly affected by dataset size and class imbalance, with larger, more balanced datasets generally leading to better results.
Key Results
- HateBERT and CyberBERT outperform vanilla BERT: On the Twitter dataset, HateBERT achieved an F1 score of 0.885 and an accuracy of 0.873, while CyberBERT achieved 0.849 F1 and 0.861 accuracy. Vanilla BERT lagged behind with 0.705 F1 and 0.692 accuracy.
- Proposed model shows significant improvement: The proposed MLP model incorporating sarcasm detection achieved the highest performance across all datasets. For instance, on the Twitter dataset, it reached an F1 score of 0.937 and an accuracy of 0.924, outperforming HateBERT by approximately 5% in both metrics.
- Sarcasm detection is the primary driver of improvement: The ablation study showed minimal improvement when only the MLP was added without the sarcasm score, confirming that the sarcasm detection component is the main contributor to the performance gains.
Limitations and Future Work
- Dataset bias: The authors acknowledge that all datasets are human-annotated, which can introduce bias based on annotator knowledge and definitions. Additionally, some datasets are dated, potentially not reflecting current cyberbullying trends.
- Generalizability: The current method is tested on social media texts, and its effectiveness on other platforms or text types is not yet established.
- Sarcasm detector accuracy on cyberbullying data: The paper does not evaluate the sarcasm classifier's accuracy specifically on the cyberbullying dataset itself. It's possible the sarcasm detector might be picking up on underlying features correlated with cyberbullying rather than pure sarcasm.
- Future work: The authors suggest exploring other NLP tasks like sentiment and emotion analysis, or intent classification, to further enhance cyberbullying detection. They also propose investigating the difference between hate speech and cyberbullying definitions for potentially improved classification. Additionally, exploring more complex deep learning architectures for sarcasm detection could be beneficial.
Summary updated