Towards Knowledge-Grounded Counter Narrative Generation for Hate Speech

Yi-Ling Chung, Serra Sinem Tekiroğlu, Marco GueriniACL 2021

TL;DR

This paper tackles the challenge of generating informative and evidence-based counter-narratives (CNs) to combat online hate speech. They introduce a novel pipeline that grounds CN generation in external knowledge, moving beyond generic responses to provide factual and persuasive arguments. This approach significantly improves the quality and relevance of counter-narratives, especially in zero-shot cross-domain scenarios.

What Problem Does This Paper Address?

Online hate speech is a pervasive problem, and while content moderation is a common approach, it has limitations regarding free speech and diverse opinions. Counter-narratives offer an alternative by directly engaging with hateful content and providing non-negative, fact-based responses. However, current automated CN generation methods often produce generic, repetitive, or even factually incorrect (hallucinated) responses. They lack the credible evidence, facts, and examples that make human-composed CNs effective. This paper aims to bridge this gap by developing a system that can automatically generate informative and knowledge-grounded counter-narratives.

Methods

The authors propose a knowledge-grounded counter-narrative generation pipeline with two main modules:

  1. Knowledge Retrieval Module: This module is responsible for finding relevant external knowledge. It involves:
  2. Knowledge Repository: A collection of unstructured text from sources like Newsroom and WikiText-103.
  3. Query Construction: This is a novel aspect. They use two strategies:
  4. Query Extraction (Qhs, Qcn, QhsUcn): Extracting keyphrases from the hate speech (HS), the counter-narrative (CN), or both. They found that using keyphrases from both HS and CN (QhsUcn) yielded the best results for retrieval.
  5. Query Generation (Qgen): Training a Transformer-based model to generate keyphrases for queries, especially useful when gold CNs are not available during testing.
  6. Knowledge Sentence Selection: Using BM25 to rank and select the most relevant sentences from retrieved articles, referred to as "silver knowledge."
  7. Counter Narrative Generation Module: This module uses pre-trained language models (LMs) fine-tuned on the retrieved knowledge and the HS. They experiment with three models:
  8. GPT-2KN: Fine-tuning GPT-2 on HS-CN pairs augmented with retrieved knowledge.
  9. GPT-2KN,MT: A multi-task version of GPT-2KN that also predicts if a response is a suitable next utterance, aiming to reduce abusive language.
  10. XNLG: A pre-trained Transformer model known for its ability to copy information, tested for its potential to incorporate retrieved knowledge.

Datasets

  • CONAN Dataset: This is the primary dataset used. It consists of 6645 English HS-CN pairs, sourced from expert NGO operators. The dataset includes original pairs, paraphrased versions, and translated pairs. For training, they used 4069/1288/1288 samples for train/dev/test.
  • Newsroom and WikiText-103: Used as the knowledge repository for retrieving relevant information.
  • Gold Knowledge Test Set: A new test set created by expert operators for evaluating the models' ability to use provided knowledge, particularly for cross-domain scenarios.

Key Insights and Ideas

  • Knowledge Infusion is Crucial: Simply fine-tuning LMs on HS-CN data is insufficient. Injecting external, up-to-date knowledge significantly improves the quality and informativeness of counter-narratives.
  • Keyphrase Generation for Retrieval: Automatically generating keyphrases (Qgen) is a viable strategy for retrieving relevant knowledge, especially when gold CNs are not available. This is a key innovation for enabling knowledge retrieval in real-world scenarios.
  • Leveraging "Silver Knowledge": The approach demonstrates that even "silver knowledge" (retrieved knowledge that isn't explicitly curated for each HS-CN pair) can be effectively used to generate informative CNs.
  • Zero-Shot Cross-Domain Robustness: Knowledge-grounded models show promising performance in generating counter-narratives for new, unseen hate targets (cross-domain) without requiring additional training, highlighting the generalizability of the approach.
  • XNLG's High Novelty is a Double-Edged Sword: While XNLG achieves high novelty and diversity scores, this is largely due to its tendency to copy entire knowledge sentences, rather than truly generating novel arguments. This points to the importance of balancing knowledge incorporation with genuine generation.

Key Results

  • Improved Performance: Knowledge-grounded models (GPT-2KN, GPT-2KN,MT, XNLG) consistently outperform baselines (TRF, GPT-2) in terms of novelty, BLEU-2, and ROUGE-L scores on the silver knowledge test set.
  • Query Strategy Matters: The QhsUcn query strategy (combining HS and CN keyphrases) provided the best retrieval results, indicating the benefit of considering both the hate speech and the desired counter-argument.
  • Human Evaluation: Human evaluations confirmed that GPT-2KN generated the most suitable CNs, while XNLG produced the most informative ones. GPT-2KN's performance was notably better than GPT-2, especially in cross-domain settings, underscoring the value of knowledge grounding.
  • Cross-Domain Success: GPT-2KN showed robust performance in cross-domain scenarios, suggesting its ability to adapt to new hate targets.

Limitations and Future Work

  • Potential for Hallucination: While the approach reduces hallucination, it's not entirely eliminated. Future work could involve an additional step to verify generated text against provided knowledge.
  • Abusive Language Risk: GPT-2 models can sometimes generate abusive language. The GPT-2KN,MT model attempts to mitigate this, but further research into controllable generation methods is suggested.
  • Query Generation Limitations: The current query generation model's performance is still limited, and future work could explore more advanced query generation techniques, including human-crafted queries or leveraging pre-trained LMs.
  • Human Evaluation is Key: The authors emphasize that for sensitive tasks like hate speech countering, human evaluation remains the gold standard, and generation systems should ideally serve as suggestion tools for human moderators.
  • Further Architectural Exploration: Investigating alternative architectures for the generation module is a potential future direction.

Summary updated

BibTeX
@inproceedings{chung2021towards,
  author = {Yi-Ling Chung and Serra Sinem Tekiroğlu and Marco Guerini},
  title = {{Towards Knowledge-Grounded Counter Narrative Generation for Hate Speech}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
}
Figure