TL;DR
This paper introduces CounterGeDi, a novel approach to generate more effective counterspeech by controlling its politeness, detoxification, and emotional tone. By leveraging Generative Discriminators (GEDI), they demonstrate significant improvements in these attributes without sacrificing the relevance of the generated text, offering a more nuanced and controllable tool for combating online hate.
What Problem Does This Paper Address?
Online hate speech is a pervasive issue, and counterspeech is a crucial strategy to combat it. While automated generation of counterspeech can assist human moderators, existing models often lack control over the generated output's specific characteristics. This means they might produce responses that are not polite enough, still contain toxic elements, or fail to convey the appropriate emotional tone needed to effectively de-escalate or counter hateful messages. The authors aim to bridge this gap by enabling fine-grained control over these desirable attributes in generated counterspeech.
Methods
The core of this work is the CounterGeDi framework, which integrates Generative Discriminators (GEDI) with a pre-trained language model, DialoGPTm. GEDI acts as a "controller" during the generation process. It's trained to distinguish between desired and undesired attributes (e.g., polite vs. impolite, toxic vs. non-toxic). During generation, GEDI guides the DialoGPTm model to produce text that aligns with the specified positive attributes.
The authors employ a three-part pipeline:
- Vanilla Counterspeech Generation Model (DialoGPTm): This model generates initial counterspeech candidates.
- GEDI Models: Multiple GEDI models are trained, each focusing on a specific attribute (politeness, detoxification, or different emotions like joy, sadness, anger, fear). These models provide a "control signal."
- Output Token Selector: This component combines the probabilities from the DialoGPTm and GEDI models, using a weighted equation to steer the generation towards the desired attributes.
They also explore single-attribute control (e.g., only politeness) and multi-attribute control (e.g., politeness, detoxification, and joy simultaneously).
Datasets
The research utilizes three publicly available datasets for training and evaluation:
- CONAN: Contains counterspeech generated by expert NGO operators.
- Reddit: Contains counterspeech generated by crowdworkers.
- Gab: Also contains counterspeech generated by crowdworkers.
For training the GEDI models, they leverage:
- A politeness dataset with 1.39 million posts, using P9 as the polite class.
- A Kaggle dataset for detoxification, using "toxic" and "non-toxic" labels.
- A large emotion dataset (416,809 posts) with seven emotions, focusing on "sadness," "joy," "fear," and "anger."
Key Insights and Ideas
- Controllability is Key: The most significant insight is that fine-grained control over counterspeech attributes is not only possible but also beneficial. Simply generating generic counterspeech is less effective than tailoring it.
- GEDI as an Effective Controller: The GEDI framework proves to be a robust method for guiding generation models towards specific attributes without significantly degrading the quality or relevance of the output.
- Complementary Attributes Enhance Effectiveness: Combining attributes like politeness, joyfulness, and detoxification can lead to even better results than single-attribute control, suggesting a synergistic effect.
- Relevance is Maintained: A crucial finding is that applying these controls does not negatively impact the relevance of the generated counterspeech to the original hate speech. In some cases, relevance even improves.
- Fluency Trade-off: While controllability is enhanced, there's a slight observed drop in fluency, which the authors acknowledge as a potential cost of adding these control mechanisms.
Key Results
- Significant Attribute Improvement: Single-attribute GEDI models improved politeness by approximately 15% and detoxification by 6%. Emotional attributes, particularly "joy," saw improvements of up to 88% on the CONAN dataset.
- Multi-Attribute Synergy: Combining attributes like politeness, detoxification, and joyfulness led to further score increases compared to single-attribute setups.
- Relevance Maintained: BLEU-2 scores, a measure of relevance, showed no significant degradation and even improved in some cases across different datasets and control strategies.
- Human Evaluation Support: Human judges also perceived a higher presence of controlled attributes in the generated counterspeech compared to baselines.
Limitations and Future Work
- Fluency Trade-off: The authors acknowledge a slight decrease in fluency as a cost of controllability. Future work could explore methods to improve fluency alongside attribute control, especially in multi-attribute setups.
- Attribute Interaction Complexity: While some attribute combinations (e.g., joy + polite + detox) showed strong performance, the interactions between multiple attributes can be complex, and further research is needed to fully understand these dynamics.
- Human-in-the-Loop: The paper emphasizes that their system is designed to assist, not fully automate, counterspeech generation, requiring human participation for optimal use.
- Further Exploration of Attributes: The authors suggest exploring other attributes and more complex combinations to further refine counterspeech generation.
Summary updated