Efficient knowledge graph to text powered by LLGM: linear latent graph model

Knowledge graph to text generation is crucial for interpreting complex structured data, yet state-of-the-art transformer models face significant computational burdens, limiting their practical deployment. This paper introduces the Linear Latent Graph Model (LLGM), a novel architecture that significantly enhances efficiency in KG-to-text generation without compromising performance. LLGM’s core innovations are three-fold: (1) a Multi-head Statistical Attention (MSA) mechanism that achieves linear O(N) complexity by replacing pairwise token interactions with efficient statistical approximations, drastically reducing the primary computational bottleneck; (2) a Graph Latent Self-Attention (GLSA) module that efficiently encodes explicit graph structures using dimension-reduced intermediate representations, preserving relational fidelity with fewer parameters; and (3) a Graph Periodicity Projector (GPP) that optimizes feed-forward networks by decomposing representations into periodic and non-periodic components, adeptly capturing both regular and unique graph patterns. Experiments on the WebNLG and EventNarrative datasets demonstrate LLGM’s significant contributions: it achieves competitive text generation quality, evidenced by a mere 0.8% BLEU-4 gap to the top model and the highest CIDEr score (4.63) on WebNLG, while requiring 20–37% fewer parameters than leading models. LLGM offers a robust and scalable solution, effectively bridging the efficiency-effectiveness gap in KG-to-text generation and enabling broader application in resource-constrained environments.