TL;DR
This paper tackles the critical issue of victim blaming on social media, specifically Twitter, by developing a computational framework to identify such language. They introduce a novel transfer learning approach that leverages language modeling to better understand the nuances of victim blaming, achieving state-of-the-art results on a custom-annotated dataset.
What Problem Does This Paper Address?
Social media platforms like Twitter have become crucial spaces for survivors of sexual assault to share their experiences and seek support, often through hashtags like #metoo. However, this increased visibility also exposes them to a harsh backlash, including victim blaming and "slut shaming." Victim blaming, where survivors are held partially or fully responsible for the assault, can have severe psychological consequences, discouraging them from reporting crimes and seeking help. Manually identifying and preventing this harmful language is an arduous task due to its complex and often ambiguous nature. This paper aims to bridge this gap by creating computational tools to detect and mitigate victim blaming online.
Methods
The authors propose a single-step transfer learning approach for classifying victim-blaming tweets. Their method combines task-specific classification with language modeling using an auxiliary loss function. This approach is designed to prevent "catastrophic forgetting" and allows the model to adapt to the specific linguistic nuances of social media, particularly Twitter.
The core of their method involves:
- LM Pretraining: A word-level language model (LM) is trained on a large corpus of English tweets. This LM consists of an embedding layer, two LSTM layers, and a linear layer.
- Transfer Learning and Auxiliary Loss: The weights from the pretrained LM are transferred to a new model, to which an additional LSTM layer is added. An auxiliary LM loss is introduced during training to incorporate the pretrained language model's contribution. The total loss is a combination of the classification loss and the auxiliary LM loss. This dual-loss approach helps the model capture both the specific task of victim blaming detection and the general linguistic patterns of Twitter.
Datasets
The paper introduces a newly annotated dataset for victim blaming detection.
- Data Collection: Tweets were collected using the Twitter API, primarily through hashtags like #metoo and #sexualharassment. Additional tweets were scraped using terms related to sexual harassment (rape, slut, whore) to increase positive samples.
- Dataset Size: The final annotated dataset contains 5,070 tweets, with 1,562 classified as positive (victim blaming) samples.
- Annotation: A two-phase annotation process was employed, guided by a social scientist. A coding instrument (Table 1) was developed to identify key indicators of victim blaming language, such as references to clothing, victim's state, sexual history, and use of loaded terms. The inter-annotator agreement was measured using Cohen's Kappa, achieving a score of 0.712.
Key Insights and Ideas
- Victim blaming is distinct: The authors argue that victim blaming, while related to hate speech and sexism, has unique characteristics and psychological underpinnings that warrant separate computational study.
- Social media's dual role: Platforms like Twitter offer a "virtual bubble" for survivors but also amplify the reach and impact of victim blaming.
- Transfer learning for nuance: Standard NLP models struggle with the informal and nuanced language of social media. Transfer learning, especially when combined with auxiliary language modeling, is crucial for capturing these subtleties.
- The importance of context: The auxiliary loss function helps the model understand the broader linguistic context, which is vital for correctly identifying sarcasm or indirect forms of victim blaming.
Key Results
The proposed transfer learning model achieved state-of-the-art performance on the custom-annotated dataset, outperforming traditional ML approaches (SVM, Naive Bayes) and various deep learning models (LSTM, TextbiRNN, TextCNN, CharCNN, fasttext).
- Accuracy: The proposed method achieved an accuracy of 0.82, significantly higher than the best baseline (fasttext at 0.78).
- Efficiency: The model is described as simpler and computationally less expensive compared to some baselines.
Limitations and Future Work
- Annotation ambiguity: The authors acknowledge that some tweets, particularly those involving self-reporting or citing previous instances of victim blaming, can be challenging to classify and may lead to errors. They suggest further research into the annotation process for such cases.
- Sarcasm and indirectness: The model sometimes struggles with sarcasm and implicit victim blaming, as seen in one of the error analysis examples.
- General sexism vs. victim blaming: Distinguishing between general sexism and specific victim blaming can be difficult, leading to potential misclassifications.
Future Work:
- Multi-label classification: Extend the work to a multi-label task that considers sub-categories of victimization (e.g., secondary, primary, gender-based).
- Weighting factors: Explore different weighting factors for the language modeling and classification losses to optimize performance for various tasks.
- Further research: Encourage more research into the computational linguistic analysis of victim blaming and the portrayal of sexual assault survivors on social media.
Summary updated