TL;DR
This paper tackles the challenge of distinguishing between hate speech and offensive language, a common source of error in automated detection systems. They propose an adversarial debiasing method that helps models better separate these two classes, leading to improved performance and reduced harmful misclassifications.
What Problem Does This Paper Address?
Online hate speech is a pervasive problem, and while significant effort has gone into developing automatic detection methods, performance remains a challenge. A key difficulty lies in the ambiguity between hate speech and offensive language. Hate speech is defined as language expressing hatred towards a targeted group based on protected attributes, while offensive language uses offensive terms but isn't directed at a specific group. Misclassifying actual hate speech as merely offensive can lead to increased harm for targeted communities. Existing three-way classification approaches (hate speech, offensive, neither) often struggle with generalization and exhibit low performance, particularly for the hate speech class. This paper aims to improve the accuracy and trustworthiness of hate speech detection models by specifically addressing the confusion between hate and offensive language.
Methods
The core contribution of this paper is an adversarial debiasing approach to separate hate speech and offensive language. The general idea of adversarial debiasing involves training two models simultaneously: a classifier that predicts the actual class (hate, offensive, neither) and an adversary that tries to predict a protected attribute. In this work, the adversary is trained to predict the "offensiveness" of a text.
The authors explore two main adversarial setups:
- Adversary predicts hate speech & offensive language jointly (Adversary_joint): Here, both hate speech and offensive language are treated as "offensive" for the adversary's task. The adversary learns to predict this combined class. By forcing the classifier to generate representations that the adversary cannot use to distinguish between hate and offensive language, the classifier is implicitly pushed to separate them.
- Adversary discriminates between hate speech and offensive language (Adversary_sep): This setup is designed to actively help the classifier. The adversary learns to model the "offensiveness" property by discriminating between hate speech/neither and offensive language. This is framed as a "supportive" role for the adversary.
They also experimented with an Adversary_swear setup, where the adversary predicts the presence of swear words, aiming to force the classifier to rely on features beyond mere profanity.
For the classifier, they experimented with BERT, DistilBERT, and LSTMs, finding BERT and DistilBERT to be more effective. For the adversary, they used Feed-Forward Networks (FFNs) and LSTMs, opting for FFNs due to their efficiency.
Datasets
The authors utilized a combination of seven datasets for their experiments, covering English, German, Arabic, and Hindi, as well as a newly created multilingual dataset.
- English:
- Davidson17 (Davidson et al., 2017)
- Founta18 (Founta et al., 2018)
- HasocEn19 (Mandl et al., 2019)
- German:
- GermEval18 (Wiegand et al., 2018)
- HasocDe19 (Mandl et al., 2019)
- Arabic:
- L-HSAB19 (Mulki et al., 2019)
- Hindi:
- HasocHin19 (Mandl et al., 2019)
- Multilingual: A combined dataset of all the above.
They also explored the impact of class imbalance by creating a "uniform dataset" where training sets were downsampled to have an equal number of examples per class.
Key Insights and Ideas
- The insidious nature of hate speech misclassification: The paper strongly emphasizes that misclassifying hate speech as merely offensive is not just a performance issue but can lead to tangible harm for targeted groups.
- Adversarial debiasing as a tool for class separation: The core idea is that by training an adversary to fail at distinguishing between hate and offensive language (or to explicitly help separate them), the main classifier is forced to learn more robust and discriminative representations.
- Context over keywords: Error analysis suggests that the adversarial models are better at understanding context and combinations of words, rather than relying solely on individual keywords or swear words, which can be misleading.
- Balancing performance: The goal is to achieve both high overall performance and a good true positive rate for hate speech, striking a balance that is crucial for real-world applicability.
- Class imbalance is a major hurdle: The paper highlights the significant impact of class imbalance and shows that addressing it (e.g., through uniform sampling) can dramatically improve the true positive rate for hate speech, even at the cost of overall accuracy.
Key Results
- The Adversary_sep setup generally performed best, showing improvements in the true positive rate (TPH) for hate speech across several datasets, often with minimal drops in overall accuracy or F1-scores.
- For the Davidson17 dataset, the Adversary_swear method improved TPH, but the Adversary_sep method did not show positive findings.
- On HasocEn19, TPH dropped, but overall performance increased significantly (over 0.2 F1 points).
- GermEval18 saw stronger performance gains, with a +5.7% increase in TPH and a +0.02 increase in F1-score with the adversarial method.
- In the multilingual setting, the adversarial debiasing offered a better TPH score with a minimal drop in performance, and improved upon the baseline model in other metrics.
- The uniform dataset approach significantly boosted the TPH for hate speech from 22.0% to 81.8% on Davidson17, though overall accuracy dropped by 16%.
Limitations and Future Work
- Generalization challenges: The authors acknowledge that their method's generalization to languages not explicitly examined might be limited, noting issues with Arabic.
- Specific problem scope: The work focuses on a specific problem (hate vs. offensive vs. neither) and specific languages, suggesting that further research is needed for broader coverage.
- Adversary_swear limitations: The Adversary_swear method is dependent on the availability of swear word dictionaries, limiting its applicability.
- Future work: The paper suggests exploring other adversarial setups and further investigating the impact of class rebalancing. The error analysis also points towards the need for models that can better understand context and word combinations.
Summary updated