UIRISC at SemEval-2023 Task 10: Explainable Detection of Online Sexism by Ensembling Fine-tuning Language Models

Tianyun Zhong, Runhui Song, Xunyuan Liu, Juelin Wang, Boya Wang, Binyang LiSemEval @ ACL 2023 · DOI

TL;DR

This paper tackles the challenge of detecting subtle and implicit online sexism, which traditional keyword-based methods struggle with. The authors propose a novel system, SEFM, that ensembles fine-tuned language models and incorporates specialized modules to improve accuracy and explainability across three distinct sexism detection subtasks. Their approach achieved strong performance, highlighting the effectiveness of ensembling and task-specific enhancements.

What Problem Does This Paper Address?

Online platforms have become breeding grounds for sexist expressions, ranging from overt abuse to more insidious forms of prejudice and stereotyping. These harmful expressions can have significant real-world consequences, impacting user experience and contributing to a toxic online environment. While previous attempts at detecting sexism often relied on keyword matching, these methods are insufficient for identifying implicit sexism, where the offensive nature of the text is conveyed through context, nuance, or subtle phrasing rather than explicit offensive terms. The SemEval-2023 Task 10, "Explainable Detection of Online Sexism (EDOS)," aims to address this by requiring systems to not only detect sexism but also provide explanations for their classifications across different granularities. This paper focuses on building a robust and explainable system for this challenging task.

Methods

The core of the proposed system, SEFM (System of Ensembling Fine-tuning Models), is built upon several key components:

  • Task-Adaptive Pre-training (TAP): Instead of directly fine-tuning pre-trained language models (PLMs) on the downstream task, the authors first adapt them to the domain of social media text. This involves further pre-training on large amounts of unlabeled data using Masked Language Model (MLM) and next sentence prediction tasks, followed by training on task-relevant unlabeled data. This process aims to imbue the models with a better understanding of the language used in online platforms.
  • Ensembling Fine-tuned PLMs: The system utilizes four different task-adaptive pre-trained language models (BERT, ALBERT, RoBERTa, and ERNIE 2.0) and fine-tunes them for the sexism detection task. The final predictions are then aggregated using a majority voting mechanism.
  • Data Augmentation: To combat the issue of imbalanced datasets, the authors employ data augmentation techniques beyond simple oversampling. These include synonym replacement, random insertion of synonyms, random word order exchange, and random word deletion, all applied with slight perturbations to avoid overfitting.
  • Specialized Modules for Subtasks:
    • Sexism Indicator for Subtask A (SIA): This module addresses potential disagreements among annotators by assigning different weights to "Gold" (consistently labeled) and "Silver" (disagreed upon) data during training.
  • Feedback to B (FB) and Fine-grained Indicator for C (FIC): For subtasks B (categorization of sexism) and C (fine-grained vector of sexism), the system incorporates feedback loops. FB uses results from subtask C to refine subtask B predictions, and FIC uses results from subtask B to enhance subtask C classification. This creates a mutually reinforcing mechanism.
  • Loss Function: The authors opt for Focal Loss, which is specifically designed to address class imbalance by down-weighting the loss for easily classified examples and focusing more on hard-to-learn samples.

Datasets

The primary dataset used is the SemEval-2023 Task 10 dataset (EDOS), which comprises 14,000 annotated instances from Gab and Reddit. This dataset is characterized by imbalanced data distribution across all three subtasks, particularly for certain fine-grained categories in subtask C. The authors also mention using a 10% development set for evaluation.

Key Insights and Ideas

  • Implicit Sexism is a Major Hurdle: The paper strongly emphasizes that traditional keyword-based methods are inadequate for detecting the nuanced and implicit forms of sexism prevalent online. This underscores the need for more sophisticated NLP techniques.
  • Task-Adaptive Pre-training Pays Off: Adapting large language models to the specific domain of social media text before fine-tuning significantly improves their performance on sexism detection tasks.
  • Ensembling for Robustness: Combining predictions from multiple fine-tuned models, even variants of the same architecture, can lead to more robust and accurate results, especially when diversity among models is considered.
  • Feedback Loops Enhance Granularity: The proposed FB and FIC modules, which allow subtasks to inform each other, demonstrate a promising approach for improving performance on more granular classification tasks like subtask C.
  • Focal Loss is Crucial for Imbalance: The choice of Focal Loss is highlighted as a key factor in effectively handling the severe class imbalance present in the dataset, leading to better performance on minority classes.

Key Results

The SEFM system achieved the following macro F1 scores:

  • Subtask A (Binary Sexism Detection): 0.8538
  • Subtask B (Category of Sexism): 0.6619
  • Subtask C (Fine-grained Vector of Sexism): 0.4641

The system ranked 12th in subtask B, demonstrating competitive performance. Ablation studies confirmed the effectiveness of the SIA, FB, and FIC modules, as well as the data augmentation techniques. The paper also noted that while ensembling BERT variants showed improvements, the lack of diversity among these models limited the potential gains from ensemble methods.

Limitations and Future Work

  • Limited Diversity in Ensembles: The authors acknowledge that ensembling exclusively BERT variants, which have similar underlying architectures, resulted in limited diversity and thus constrained the performance gains from ensembling.
  • Challenges with Nuance and Ambiguity: The paper points out that colloquialisms, slang, spoken words, metaphors, and ambiguous sentences in social media text pose significant challenges for models.
  • Future Work: The authors plan to further improve their system by incorporating prompt modules and exploring fusion label vectors to enhance the performance of online sexism detection. They also suggest exploring more diverse model architectures for ensembling to maximize performance gains.

Summary updated

BibTeX
@inproceedings{zhong2023uirisc,
  author = {Tianyun Zhong and Runhui Song and Xunyuan Liu and Juelin Wang and Boya Wang and Binyang Li},
  title = {{UIRISC at SemEval-2023 Task 10: Explainable Detection of Online Sexism by Ensembling Fine-tuning Language Models}},
  booktitle = {SemEval @ ACL},
  year = {2023},
  doi = {10.18653/v1/2023.semeval-1.287},
}
Figure