DynaSemble: Dynamic Ensembling of Textual and Structure-Based Models for Knowledge Graph Completion
Ananjan Nandi, Navdeep Kaur, Parag Singla, Mausam
TL;DR
This work tackles Knowledge Graph Completion by combining complementary signals from textual and structure-based models through DynaSemble, a dynamic, query-dependent ensembling framework. It defines the ensemble score as $E(h,r,t) = \sum_{i=1}^{k} w_i(q)\,M_i(h,r,t)$, normalizes model outputs with min–max scaling, and learns per-model weights $w_i(q)$ via lightweight MLPs that take mean–variance features of score distributions as input. Across WN18RR, FB15k-237, and CoDex-M, DynaSemble achieves state-of-the-art results, with notable gains over individual models and static ensembling, and generalizes to RotE embeddings; analyses show improved handling of both reachable and unreachable tails. The approach provides a practical, extensible baseline for fusing heterogeneous KGC signals, with code available for reproducibility and future extensions into temporal and multilingual KGC settings.
Abstract
We consider two popular approaches to Knowledge Graph Completion (KGC): textual models that rely on textual entity descriptions, and structure-based models that exploit the connectivity structure of the Knowledge Graph (KG). Preliminary experiments show that these approaches have complementary strengths: structure-based models perform exceptionally well when the gold answer is easily reachable from the query head in the KG, while textual models exploit descriptions to give good performance even when the gold answer is not easily reachable. In response, we propose DynaSemble, a novel method for learning query-dependent ensemble weights to combine these approaches by using the distributions of scores assigned by the models in the ensemble to all candidate entities. DynaSemble achieves state-of-the-art results on three standard KGC datasets, with up to 6.8 pt MRR and 8.3 pt Hits@1 gains over the best baseline model for the WN18RR dataset.