PathISE: Learning Informative Path Supervision for KGQA
PathISE is a cutting-edge approach for Knowledge Graph Question Answering (KGQA) that focuses on obtaining high-quality supervision through answer-level labels. It includes a streamlined transformer-based estimator that assesses the usefulness of relation paths, generating pseudo supervision at the path level. This supervision is then transformed into a path generator for large language models (LLMs), ensuring the paths are grounded in the knowledge graph (KG) to provide clear evidence for reasoning in answers. This technique addresses the challenge of obtaining intermediate supervision signals, like relevant paths or subgraphs, which often demand considerable resources. Ultimately, PathISE aims to improve retrieval-augmented generation by merging structured knowledge from KGs into LLMs.
Key facts
- PathISE is a framework for Knowledge Graph Question Answering (KGQA).
- It learns intermediate supervision from answer-level labels.
- Uses a lightweight transformer-based estimator to evaluate relation path informativeness.
- Constructs pseudo path-level supervision.
- Distills supervision into an LLM path generator.
- Paths are grounded in the KG for compact evidence.
- Addresses the high cost of obtaining intermediate supervision signals.
- Aims to improve retrieval-augmented generation with KG grounding.
Entities
—