New Embedding Methods Improve Horn Logic Reasoning
A new arXiv paper (2605.20467) introduces techniques for training neural networks to rank logical reasoner choices, enhancing search efficiency. The key innovation is creating better numeric embeddings of logical statements using triplet loss. Three approaches are proposed: generating anchors with repeated terms, balancing easy/medium/hard examples during training, and periodically emphasizing hardest examples. Experiments compare embeddings across knowledge bases to identify optimal characteristics.
Key facts
- arXiv paper 2605.20467
- Published on arXiv
- Focuses on embeddings for Horn logic reasoning
- Uses triplet loss for training
- Three new techniques introduced
- Anchors with repeated terms
- Balanced positive/negative examples
- Periodic emphasis on hardest examples
- Experiments compare embeddings across knowledge bases
Entities
Institutions
- arXiv