iREL at SemEval-2023 Task 10: Multi-level Training for Explainable Detection of Online Sexism

Nirmal Manoj, Sagar Joshi, Ankita Maity, Vasudeva VarmaACL 2023

TL;DR

This paper tackles the challenge of explainable online sexism detection by proposing a multi-level, hierarchical classification approach. They leverage domain-adaptive pretraining and a focal loss objective to improve performance on imbalanced datasets, demonstrating significant gains, particularly with the RoBERTa-large model.

What Problem Does This Paper Address?

Online sexism is a pervasive issue, and while automated detection systems exist, they often lack explainability. This means users and developers don't fully understand why a piece of content is flagged as sexist, hindering trust and effective intervention. The SemEval-2023 Task 10 aims to address this by introducing a hierarchical taxonomy of sexist content, moving beyond simple binary classification to more granular categories. This granular approach is crucial because the motivations behind sexist remarks can vary widely, requiring nuanced detection for appropriate handling. The paper focuses on building models that can not only detect sexism but also categorize it in a way that offers more insight into its nature.

Methods

The authors employed a multi-pronged approach to tackle the hierarchical sexism detection task:

  • Basic Finetuning: This is the standard approach of taking a pretrained transformer model (like BERT or RoBERTa) and fine-tuning it on the specific downstream task.
  • Focal Loss: To address the significant class imbalances present in the finer-grained tasks (Tasks B and C), they utilized Focal Loss. This loss function downweights easy-to-classify examples and focuses the model's attention on harder, misclassified examples, which is particularly helpful for minority classes.
  • Domain-Adaptive Pretraining (DAPT): Recognizing that online social media language differs from general text, they pre-trained their models further on a large corpus of unlabeled data from Reddit and Gab (the same platforms where the task data originated). This helps the models adapt to the specific linguistic nuances, slang, and informalities common in these environments.
  • Multi-level Training: This is a key novel contribution. They exploit the hierarchical nature of the task by using the best performing checkpoint from a coarser task (e.g., Task A) as the starting point for training the next, more granular task (e.g., Task B). This allows knowledge transfer from simpler to more complex classification levels.

Datasets

The primary dataset used was the SemEval-2023 Task 10 dataset, which includes labeled instances for three subtasks:

  • Task A: Binary classification (sexist vs. non-sexist).
  • Task B: Classification into 4 categories of sexism for sexist posts.
  • Task C: Classification into 11 "fine-grained vectors of sexism" for sexist posts.

Additionally, two unlabeled datasets from Reddit and Gab, each containing 2 million comments, were used for Domain-Adaptive Pretraining.

Key Insights and Ideas

  • Hierarchy Matters: The multi-level training approach proved beneficial, especially for Task B, demonstrating that learning from a broader category can significantly improve performance on more specific ones.
  • Domain Adaptation is Crucial: Pretraining on social media data (Reddit/Gab) significantly boosted performance, highlighting the importance of adapting models to the specific domain of online discourse.
  • Focal Loss for Imbalance: The focal loss objective was particularly effective in Task C, where class imbalances were most severe, leading to notable performance improvements.
  • RoBERTa-Large Dominance: The RoBERTa-large model consistently showed superior performance across various training strategies and tasks, underscoring its power for this type of nuanced text classification.

Key Results

The paper reports strong performance on the validation sets:

  • Task A: The best macro-F1 score achieved was 85.93% using RoBERTa-large with domain-adaptive pretraining.
  • Task B: The best macro-F1 score was 69.96%, also achieved by RoBERTa-large, utilizing multi-level training with pretraining.
  • Task C: The best macro-F1 score was 54.62%, obtained with RoBERTa-large using basic finetuning with focal loss.

The authors also note that domain-adaptive pretraining generally improved models on Task A, while multi-level training was most effective for Task B. Focal loss showed a pronounced benefit in Task C due to its high class imbalance.

Limitations and Future Work

  • Model Specificity: While RoBERTa-large performed exceptionally well, the paper doesn't extensively explore if other architectures could achieve similar results with different training strategies.
  • Explainability Depth: While the task is "explainable detection," the paper focuses more on classification performance. The actual "explainability" aspect (e.g., generating justifications) isn't detailed in this work, which is a natural next step.
  • Generalizability: While DAPT improved domain-specific performance, further work could explore broader generalization across different social media platforms or even to other forms of online toxicity.
  • Future Work: The authors suggest exploring more sophisticated multi-level training strategies and investigating how to further enhance the explainability of the detection models.

Summary updated

BibTeX
@inproceedings{manoj2023irel,
  author = {Nirmal Manoj and Sagar Joshi and Ankita Maity and Vasudeva Varma},
  title = {{iREL at SemEval-2023 Task 10: Multi-level Training for Explainable Detection of Online Sexism}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure