TL;DR
This paper tackles the challenge of detecting online sexism with a hierarchical, multi-task learning approach. By pre-training large language models on a massive unlabeled dataset and then fine-tuning them within a multi-task framework, the authors achieved state-of-the-art results on a challenging sexism detection benchmark, demonstrating the power of combining these techniques for nuanced and explainable AI.
What Problem Does This Paper Address?
Online sexism is a pervasive issue with harmful consequences. While automated tools exist to detect it, they often provide only generic, high-level classifications without explaining why content is flagged. This lack of interpretability hinders effective moderation and user understanding. SemEval-2023 Task 10 specifically aims to develop more accurate and explainable models for online sexism detection, addressing a critical gap in current research. The task involves a hierarchical classification system, making it more complex than simple binary detection.
Methods
The core of the system is the MT-DNN (Multi-Task Deep Neural Network) framework, a neural network architecture designed to handle multiple related NLP tasks simultaneously. This allows the model to share knowledge and learn from the interdependencies between tasks, leading to improved generalization.
The authors employed a two-pronged strategy:
- Domain Pre-training: They took pre-trained models, specifically RoBERTa-large and DeBERTa-v3-large, and further pre-trained them on a substantial dataset of 2 million unlabeled posts from Gab and Reddit. This step helps the models adapt to the specific language and nuances of online discourse relevant to sexism.
- Multi-Task Learning: The pre-trained models were then integrated into the MT-DNN framework. This allowed them to jointly train on the three hierarchical sub-tasks of the SemEval task:
- Sub-task A: Binary classification (sexist vs. not sexist).
- Sub-task B: Categorization of sexism (threats, derogation, animosity, prejudiced discussions).
- Sub-task C: Fine-grained vector of sexism (11 categories).
Distinct "head" layers were designed for each sub-task within the MT-DNN architecture.
Datasets
- Training/Development/Test Data: 20,000 English posts collected from Reddit and Gab, split with a 0.7:0.1:0.2 ratio for training, development, and final evaluation, respectively.
- Unlabeled Data: 2 million unlabeled posts from Gab and Reddit used for domain pre-training.
The paper notes that the class distributions for all sub-tasks are imbalanced, with "not sexist" being the majority class in Sub-task A, "derogation" in Sub-task B, and "descriptive attacks" in Sub-task C.
Key Insights and Ideas
- Pre-training on Domain-Specific Unlabeled Data is Crucial: Further pre-training large language models on relevant unlabeled data (from platforms like Gab and Reddit) significantly boosted performance, especially for sub-tasks A and C. This highlights the importance of adapting general-purpose models to the specific domain of online discourse.
- Multi-Task Learning Leverages Task Interdependencies: The MT-DNN framework effectively utilized the relationships between the hierarchical sub-tasks. This allowed the model to learn more robust representations and improve performance across all tasks compared to training them in isolation.
- Hierarchical Structure Benefits from Shared Learning: The hierarchical nature of the task (binary classification first, then categorization) naturally lends itself to multi-task learning, where knowledge gained from the broader task can inform the more specific ones.
- Epoch Selection Matters: The paper found that simply increasing pre-training epochs didn't always lead to better results. Optimal epoch counts (e.g., 7 for RoBERTa, 9 for DeBERTa on sub-task A) were identified through experimentation, suggesting a need for careful hyperparameter tuning.
Key Results
The system achieved impressive results:
- Sub-task A (Binary Sexism Detection): A macro F1-score of 0.8746, ranking 1st on the leaderboard.
- Sub-task B (Category of Sexism): A macro F1-score of 0.7094.
- Sub-task C (Fine-grained Vector of Sexism): A macro F1-score of 0.5308.
- The system ranked 5th on sub-tasks B and C.
- A "Multi-model Fusion" approach for Sub-task A achieved an F1 score of 0.8751.
Limitations and Future Work
- Class Imbalance: The paper acknowledges the imbalanced nature of the datasets across all sub-tasks, which can pose challenges for model training and evaluation.
- Hyperparameter Sensitivity: While the paper details their chosen hyperparameters, the optimal settings might vary for different datasets or model architectures.
- Explainability: While the task aims for explainability, the paper focuses primarily on achieving high F1 scores. The specific mechanisms for generating explanations are not detailed in this excerpt.
Future work could explore more advanced explainability techniques to complement the high performance achieved, and potentially investigate methods to further mitigate the impact of class imbalance.
Summary updated