TL;DR
This paper introduces a novel multitask learning framework (MFAE) that jointly tackles abuse detection and emotion classification. By leveraging emotion classification as an auxiliary task, MFAE significantly improves abuse detection performance, especially on datasets with limited labeled abuse data.
What Problem Does This Paper Address?
Online social media has become a double-edged sword: while it connects people, it also amplifies abusive language. Automatically detecting this abuse is crucial for mitigating its severe psychological impact. Existing methods often focus solely on linguistic features, neglecting the crucial role of user emotions. This paper addresses this gap by proposing a framework that integrates emotion classification to enhance abuse detection, hypothesizing that understanding a user's emotional state can lead to more accurate abuse identification.
Methods
The core of MFAE is a multitask learning architecture built upon a pre-trained BERT model. The process involves:
- Emotional Label Generation: Since many abuse detection datasets lack explicit emotion labels, the authors employ a pre-trained sentiment analysis model (SKEP) to automatically generate soft emotion labels. This approach avoids costly manual annotation and addresses the issue of imbalanced datasets by using probability distributions rather than hard labels.
- Encoder: A fine-tuned BERT model serves as the encoder to generate rich, contextualized sentence representations. The first 8 BERT layers are frozen, while the last 4 are updated to balance convergence speed and representation stability.
- Decoder: Two distinct decoders are used for the two tasks:
- Emotion Classification Decoder: A BiLSTM network processes BERT outputs to capture contextualized representations, followed by max-pooling and a fully connected layer with softmax to predict emotion probabilities.
- Abuse Detection Decoder: A TextCNN model is employed to extract local context features, leveraging dynamic max-pooling to capture sentence-level representations.
- Cross-Attention (CA) Component: A novel cross-attention mechanism is introduced to allow the emotion classification decoder to influence the abuse detection decoder. This component helps the abuse detection task focus on relevant emotional cues, further enhancing its performance.
Datasets
The framework was evaluated on five publicly available datasets:
- HatEval: A Twitter-based hate speech dataset focusing on hate speech against immigrants and women.
- Davids: A Twitter-based online abusive language detection dataset with categories for Hate, Offensive, and Neither.
- OffEval: A dataset for offensive language identification.
- FNUC: A lightweight hate speech detection dataset from Fox News discussion threads.
- StormW: A hate language detection dataset from the Stormfront forum.
The authors preprocessed all datasets by removing URLs, converting text to lowercase, and padding/truncating sequences to a uniform length of 50.
Key Insights and Ideas
- Emotion as a Powerful Auxiliary Task: The study strongly suggests that incorporating emotion classification as an auxiliary task significantly boosts abuse detection performance, even outperforming models that solely rely on advanced pre-trained language models like BERT.
- Automated Label Generation is Viable: The authors demonstrate that using pre-trained sentiment analysis models to generate emotion labels is a practical and effective alternative to expensive manual annotation, especially for datasets lacking such labels.
- Cross-Attention Bridges the Gap: The proposed cross-attention mechanism is crucial for effectively transferring knowledge from the emotion classification task to the abuse detection task, proving indispensable for achieving optimal results.
- Multitask Learning Enhances Robustness: The framework shows improved performance across various datasets, indicating that multitask learning can help mitigate overfitting, particularly on smaller datasets like FNUC.
Key Results
The proposed MFAE framework achieved state-of-the-art results across all five tested datasets, outperforming various baseline models, including word-level, character-level, RNN-based, CNN-based, Transformer-based, and even BERT-based approaches.
- Weighted F1 Score: MFAE achieved the highest average weighted F1 score of 82.05%, outperforming the next best model (BERT) by approximately 1.7%.
- Accuracy Score: Similarly, MFAE achieved the highest average accuracy score of 82.316%, again surpassing BERT.
- Ablation Studies: Removing any of the core components (decoder, cross-attention, BiLSTM, or TextCNN) led to a decrease in performance, underscoring the importance of each element in the MFAE architecture.
- Parameter Sensitivity: The weighting parameter ($\lambda$) for the auxiliary task's loss demonstrated that a higher weight (typically greater than 0.5) for emotion classification leads to better abuse detection performance, highlighting the task's significant positive impact.
Limitations and Future Work
- Noise in Pseudo-Labels: The authors acknowledge that the automatically generated emotion labels, derived from a pre-trained model, may contain noise, which can sometimes lead to unreliable emotion classification and potentially mislead abuse detection.
- Simple Emotion Dichotomy: The current framework uses a binary emotion classification (positive/negative). Future work aims to expand this to multi-class emotion classification.
- Combining with Other Pre-trained Models: The authors plan to explore combining their framework with other pre-trained models beyond BERT.
- Improving Pseudo-Labeling: Future research will focus on developing more robust methods to reduce noise in pseudo-labels or explore alternative derivation techniques.
Summary updated