Fuzzy knowledge graph embedding (KGE) aims to represent uncertain facts for reliable reasoning and decision-making. While recent studies have begun to explore fine-grained fuzzy KGs that associate uncertainty with individual elements, existing approaches remain limited by (i) unreliable neighborhood aggregation that may amplify low-confidence noise, (ii) entangled semantic and reliability signals that obscure where uncertainty originates and hinder interpretability, and (iii) restricted generality due to task- or metadata-specific designs. To address these issues, we propose FGF-GAT, a general encoder-decoder framework for fine-grained fuzzy KGE. On the encoder side, FGF-GAT employs a learnable neuro-fuzzy reasoning module for fine-grained fuzziness modeling, where fixed element-level memberships are transformed into reliability-aware signals through an ANFIS-inspired fuzzy parameterization and a lightweight TSK-style rule reasoning layer. These memberships are then injected into a rule-constrained fuzzy graph attention encoder for reliability-modulated neighborhood aggregation, thereby suppressing uncertainty-induced noise accumulation. On the decoder side, we develop a pluggable membership injection strategy and demonstrate that the framework is decoder-agnostic by instantiating a basic Fuzzy-TransE scorer and two stronger variants with DistMult and RotatE. Extensive experiments on four benchmarks under both entity prediction and relation prediction, together with ablation and sensitivity studies, consistently verify the effectiveness, robustness, and adaptability of the proposed framework.
Li et al. (Thu,) studied this question.