Knowledge‑Enhanced Graph Semantic Understanding and the ERNIESage Framework

The presentation outlines the concept of knowledge‑enhanced graph semantic understanding, combining pre‑trained language models (ERNIE 1.0, 2.0) with graph neural networks (GraphSAGE) to capture both textual semantics and external structured knowledge.

ERNIESage extends this idea through several variants: ERNIESage‑Node attaches ERNIE to graph nodes, ERNIESage‑Edge moves the ERNIE module to edges for token‑level interaction, ERNIESage‑1‑Neighbor samples and aggregates first‑order neighbors, and ERNIESage‑N‑Neighbor scales the approach to multiple hops, each increasing computational complexity.

The framework also distinguishes explicit knowledge injection—adding external knowledge graphs as neighbor nodes—from implicit injection, where knowledge is absorbed during pre‑training via masked token strategies and transformer‑based fusion, allowing the model to retain knowledge without extra online resources.

Industrial case studies demonstrate the effectiveness of ERNIESage: (1) keyword‑trigger generation for search queries, where heterogeneous query graphs are modeled to improve retrieval, and (2) map POI ranking, where lightweight ERNIE‑Tiny handles online queries while a richer ERNIESage model encodes POI relationships offline for fast matching.

Additional Q&A highlights multi‑modal extensions (e.g., ERNIE‑Vil for image‑text), the rationale for masking tokens before transformer processing, and practical considerations for graph node selection and query length constraints.

In summary, integrating knowledge graphs with semantic models via ERNIESage enhances both explicit and implicit knowledge utilization, yielding measurable gains across search, recommendation, and map services.