Zero-Shot NLI Stance Detection for Knowledge Graph Edge Classification: Validating the Inference Level

Key Points

• The aim is to validate inference-level edge classification methods for stance detection in knowledge graphs.
• Used mDeBERTa-v3 for zero-shot natural language inference (NLI) to classify edges as CONFIRMS or CONTRADICTS.
• Compared performance with Claude Haiku and fine-tuned BERT baselines.
• Performed evaluations across three different domains and two languages.
• Demonstrated effective edge classification using mDeBERTa-v3 in zero-shot scenarios.
• Achieved improved accuracy over BERT and Claude Haiku in stance detection tasks.
• Showed consistent results across both languages and domains.