TL;DR
This paper tackles the crucial problem of automatically detecting and categorizing online sexism across three hierarchical tasks, using both CNN-BiLSTM and GPT-2 models. They found GPT-2 to be a more effective model, especially for the more granular tasks, and developed a web application called 'PREDHATE!' for practical application.
What Problem Does This Paper Address?
Online sexism is a pervasive and harmful issue that contributes to hostile environments and social inequities. While its impact is significant, automatically identifying and understanding sexist content online remains a challenge. This paper addresses the need for robust models that can not only detect sexism but also categorize its nuances and provide explanations, thereby empowering users and moderators to combat this problem more effectively. The SemEval-2023 Task 10 provided a benchmark for this, focusing on a hierarchical approach to sexism detection.
Methods
The authors employed two primary deep learning architectures:
- CNN-BiLSTM: This model combines Convolutional Neural Networks (CNNs) for feature extraction (identifying n-grams and word embeddings) with Bidirectional Long Short-Term Memory (BiLSTM) layers for capturing long-range dependencies in sequential text data. This hybrid approach aims to leverage the strengths of both for text classification.
- GPT-2: A pre-trained Transformer-based language model, GPT-2 was fine-tuned for the sexism detection tasks. Its ability to understand context and generate coherent text makes it a powerful tool for natural language understanding tasks.
The problem was broken down into three hierarchical subtasks:
- Task A (Binary Sexism Detection): Classifying text as either sexist or not sexist.
- Task B (Category of Sexism): For sexist posts, categorizing them into one of four classes: Threats, Derogation, Animosity, and Prejudiced Discussions.
- Task C (Fine-grained Vector of Sexism): Further classifying sexist posts into one of eleven fine-grained vectors, detailing specific types of sexist expression.
Datasets
The primary dataset used is from SemEval-2023 Task 10, consisting of 20,000 labeled entries. This dataset was sourced from Gab and Reddit. The data was annotated by three professional annotators, with disagreements resolved by an expert. The dataset was split into training, development, and test sets. Additionally, two unlabeled datasets from Gab and Reddit, each with one million entries, were made available to facilitate training.
Key Insights and Ideas
- Hierarchical Approach is Key: Breaking down sexism detection into binary, categorical, and fine-grained levels allows for a more nuanced understanding and potentially more effective interventions.
- GPT-2's Superiority: The study strongly suggests that GPT-2, a pre-trained language model, significantly outperforms the CNN-BiLSTM model, particularly for the more complex and granular tasks (B and C). This highlights the power of large pre-trained models for fine-grained text analysis.
- Explainability Matters: The task's emphasis on explainability (though not explicitly detailed in the methods section of this paper) is crucial for building trust and enabling users and moderators to understand why a piece of text is flagged as sexist.
- Practical Application: The development of the 'PREDHATE!' web application demonstrates a commitment to making these advanced models accessible and useful in real-world scenarios.
Key Results
- Task A Performance: GPT-2 achieved a Macro F1 score of 0.75 for Task A, outperforming the CNN-BiLSTM model which scored 0.73.
- Task B and C Performance: GPT-2's performance on Tasks B and C was lower than Task A, with scores of 0.40 and 0.20 respectively. This indicates the increased difficulty of these more granular classification tasks.
- Model Comparison: The paper explicitly states that GPT-2 performed better on Task A, leading them to use GPT-2 for all three tasks.
- Leaderboard Comparison: The paper includes a comparison with the top scorer on SemEval-2023 Task 10, showing their GPT-2 model achieved a Macro F1 of 0.74 for Task A, 0.41 for Task B, and 0.20 for Task C, placing them competitively.
Limitations and Future Work
- Bias in Predictions: The authors acknowledge that their model, while performing as expected, may exhibit bias and not produce accurate answers for certain groups, leading to poorer performance.
- Data Limitations: The paper mentions that for some labels in Task C, the classifier was unable to predict samples, likely due to the distribution of data in the test set.
- Future Work: The authors hope their work will stimulate further research into automated sexism identification and categorization, and that their dataset and experimental setups will serve as a valuable resource for the community. They aim for their work to contribute to creating more welcoming and secure online environments.
Summary updated