flipz357/S3BERT

Partitions sentence embeddings into interpretable sub-embeddings aligned with custom semantic metrics (e.g., coreference, negation, semantic roles), enabling aspect-specific similarity scoring without inference-time metric computation. Built on transformer-based sentence encoders (sentence-transformers) with a training procedure that routes semantic information into designated feature dimensions while maintaining overall embedding quality through consistency constraints. Includes pre-trained models based on MPNet and MiniLM architectures with evaluation on semantic similarity benchmarks and AMR-derived aspect metrics.