Core Concepts
Counterspeech represents a promising strategy to mitigate the spread of online hate, and NLP can provide the tools to make it scalable. This paper offers a comprehensive guide on how to approach counterspeech research from an NLP perspective, covering task design, data selection, and evaluation.
Abstract
This paper provides a thorough review of 43 NLP studies on counterspeech, structured as a step-by-step guide for researchers approaching this topic.
The authors first frame the concept of counterspeech, distinguishing it from related tasks like hope speech and counter-argumentation. They then outline three key steps in counterspeech research:
- Task Design:
- Counterspeech classification (detecting counterspeech, identifying strategies, classifying users)
- Counterspeech selection (retrieving appropriate responses from a pool)
- Counterspeech generation (incorporating knowledge, personality, and stylistic aspects)
- Data Selection:
- Discussing the pros and cons of different data collection methods (crawling, crowdsourcing, nichesourcing, hybrid, and fully automated)
- Describing characteristics of available counterspeech datasets, including interaction type, target of hate, language, and additional annotations
- Evaluation:
- Outlining metrics for assessing counterspeech classification and generation, including both automatic and human evaluation approaches
- Highlighting the importance of user-oriented evaluation to capture the subjective nature of effective counterspeech
Finally, the authors identify key open challenges in counterspeech research, such as addressing language and cultural specificity, handling implicit hate, mitigating hallucinations in generation, and accounting for biases in data collection.
The paper emphasizes the need for responsible and ethical practices when conducting counterspeech research, given the potential social consequences of this work.