SUTNLP at SemEval-2023 Task 10: RLAT-Transformer for explainable online sexism detection

Hamed Hematian Hemati, Sayed Hesam Alavian, Hamid Beigy, Hossein SametiACL 2023

TL;DR

This paper tackles the challenge of detecting and explaining online sexism by proposing a novel "Random Layer Adversarial Training" (RLAT) method. By integrating RLAT with powerful pre-trained language models like RoBERTa and DeBERTa, the authors significantly improve performance across multiple sexism detection subtasks, outperforming baselines and offering a more robust approach to identifying harmful online content.

What Problem Does This Paper Address?

Online sexism, characterized by prejudice, stereotyping, or discrimination based on gender, is a pervasive and harmful issue. While automated systems for detecting sexism exist, they often lack interpretability, making it difficult for users and moderators to understand why a piece of content is flagged. This hinders the creation of truly fair and inclusive online environments. The SemEval-2023 Task 10, "Explainable Detection of Online Sexism," aims to address this by developing systems that not only detect sexism but also provide explanations. This paper focuses on building such systems, particularly for the English language, and addresses the inherent difficulties in defining and detecting sexism, especially in its more nuanced forms.

Methods

The core of this work lies in the proposed Random Layer Adversarial Training (RLAT) method. Unlike traditional adversarial training that often focuses on the embedding layer, RLAT randomly selects a dense layer or the embedding layer of a transformer model and applies adversarial perturbations. This approach aims to improve the model's robustness and generalization by exposing it to a wider range of challenging inputs.

The authors also leverage:

  • Large Pre-trained Language Models (LLMs): RoBERTa-Large and DeBERTa-v3-Large are used as the foundational models for their text classification tasks. These models are known for their strong performance on various NLP tasks.
  • Virtual Adversarial Training (VAT): This technique is used to further enhance the transformer's generalization capabilities, particularly in semi-supervised settings, by minimizing the distance between outputs of original and perturbed inputs without relying on label information.
  • Contrastive Learning (CON): Employed to boost the discrimination power of the model by ensuring that samples with similar labels are represented closely in the embedding space, while dissimilar samples are pushed further apart.

For Subtask A, all three techniques (RLAT, VAT, and CON) are combined with the base transformer. For Subtasks B and C, RLAT is combined with cross-entropy loss.

Datasets

The primary dataset used in this work is the EDOS (Explainable Detection of Online Sexism) dataset, collected from Reddit and Gab. This dataset consists of approximately 20,000 samples, each tagged by three annotators. The dataset is structured into three hierarchical subtasks:

  • Subtask A (Binary Sexism Detection): Classifies comments as either sexist or not sexist.
  • Subtask B (Category of Sexism): A four-class classification for sexist posts, categorizing them into threats, derogation, animosity, and prejudiced discussions.
  • Subtask C (Fine-grained Vector of Sexism): A more granular classification into 11 distinct categories of sexism.

The paper notes that Subtasks B and C present a significant challenge due to class imbalance, with ratios of the most frequent to least frequent class being approximately 5 and 14, respectively.

Key Insights and Ideas

  • Randomness is Key for Adversarial Training: The most significant insight is that randomly selecting layers for adversarial training (RLAT) proves more effective than traditional methods that focus solely on the embedding layer. This suggests that perturbing different parts of the network can lead to more robust and generalized models.
  • RLAT's Superiority Over Conventional AT: The paper provides strong evidence that RLAT significantly outperforms conventional adversarial training (AT) applied to the embedding layer, especially for Subtask C, where it yields a substantial gain in Macro-F1 score.
  • Combined Techniques Yield Best Results: For Subtask A, combining RLAT, VAT, and contrastive learning with the base transformer leads to the highest performance, demonstrating the synergistic benefits of these approaches.
  • Data Imbalance is a Major Hurdle: The authors explicitly highlight the challenges posed by class imbalance in Subtasks B and C, indicating that future work might need to focus on more advanced techniques to address this.

Key Results

The proposed RLAT-Transformer model achieved impressive results across all subtasks, outperforming baseline methods:

  • Subtask A: The team achieved a macro-F1 score of 84.45, outperforming baselines. The full model (ROBERTa-Large + RLAT + VAT + CON) achieved a macro-F1 of 85.45.
  • Subtask B: The DeBERTa-v3-Large + RLAT model achieved a macro-F1 of 67.78, a significant improvement over the baseline (66.70).
  • Subtask C: The DeBERTa-v3-Large + RLAT model achieved a macro-F1 of 52.52, a substantial gain over the baseline (45.40). This represents a +7.12% improvement, highlighting the effectiveness of RLAT for fine-grained sexism detection.

The paper also shows that RLAT contributes most to performance boosts, with its exclusion leading to the largest drops in scores.

Limitations and Future Work

The authors acknowledge the inherent difficulty in defining and detecting sexism, especially in its more subtle forms. The class imbalance in Subtasks B and C remains a challenge.

Future work suggested includes:

  • Further Analysis of RLAT: Investigating how RLAT performs on other NLP tasks and exploring ways to further optimize its application.
  • Addressing Nuances of Sexism: Developing models that can better capture the complex and often implicit nature of online sexism.
  • Explainability Enhancements: While the task focuses on explainable detection, the paper primarily reports on detection performance. Future work could delve deeper into generating and evaluating the explanations themselves.

Summary updated

BibTeX
@inproceedings{hemati2023sutnlp,
  author = {Hamed Hematian Hemati and Sayed Hesam Alavian and Hamid Beigy and Hossein Sameti},
  title = {{SUTNLP at SemEval-2023 Task 10: RLAT-Transformer for explainable online sexism detection}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2023},
}
Figure