Knowledge Graph Completion Methods

Updated 31 December 2025

Knowledge Graph Completion is a computational paradigm that infers missing entities and relations in structured graphs by leveraging both structural and semantic context.
Approaches range from classical embedding techniques to transformer and LLM-based models, with innovations in context fusion and adaptive sampling enhancing performance.
Empirical studies demonstrate improved accuracy and scalability, with significant gains in metrics like MRR and Hits@K across diverse benchmark datasets.

Knowledge Graph Completion (KGC) is a computational paradigm devoted to inferring missing links—typically entities or relations—in structured knowledge graphs (KGs). KGC systems address the intrinsic incompleteness of KGs, which impairs downstream applications such as search, reasoning, and recommendation. The field is characterized by a wide repertoire of architectures and learning principles, ranging from classical embedding-based models to recent generative transformer and LLM-based systems. Major advances have focused on exploiting richer forms of structural and semantic context, mitigating bottlenecks in scalability and data imbalance, and enhancing prediction accuracy for long-tail entities, unseen relations, and multilingual graphs.

1. Methodological Taxonomy and Core Challenges

KGC approaches fall into several principal categories: (a) structure-based embedding methods, (b) text-based and LM-based systems, (c) contextual and multi-perspective augmentations, and (d) generative models employing seq2seq LMs.

Structure-based methods such as TransE, DistMult, ComplEx, RotatE, and TuckER learn dense, low-dimensional vector representations for entities and relations. These models typically score each possible triple $(h, r, t)$ using additive or bilinear functions and optimize with pairwise ranking or contrastive losses. However, classical embedding methods are limited by poor generalization to rare or unseen entities and exhibit degree bias—low-degree nodes are systematically under-predicted (Shomer et al., 2023).

Text-based models (e.g., KG-BERT, SimKGC) encode entities and relations using their natural language descriptions via pre-trained LLMs. While this alleviates some cold-start issues, it introduces computational costs and semantic mismatch for entities lacking robust descriptions (Gul et al., 2024, Gul et al., 5 Mar 2025). Recent embedding models incorporate contextual BERT encodings of neighboring nodes (head context, relation context), removing dependency on textual descriptions and negative sampling (see "A Contextualized BERT model for Knowledge Graph Completion" (Gul et al., 2024), "MuCo-KGC" (Gul et al., 5 Mar 2025)).

Generative models rephrase KGC as a sequence-generation problem, enabling direct decoding of entity names. Transformer encoder-decoder architectures (T5, KGT5) have demonstrated efficacy especially when enriched with contextual subgraph information (Chen et al., 29 Mar 2025, Chepurova et al., 2023). These systems often surpass structure-based methods in both predictive accuracy and scalability, particularly when careful context selection and fusion strategies are applied.

Advanced augmentations include multi-level shared knowledge distillation (Shan et al., 2024), collective completion across multiple KGs via mutual knowledge distillation (Zhang et al., 2023), and the use of common sense–enhanced modules for concept-based negative sampling and dual scoring (Niu et al., 2024). Sequence models leveraging frozen LLMs via intermediate-layer probing have further reduced training overhead by orders of magnitude (Xue et al., 2024).

2. Contextual Input Construction and Representation Fusion

Recent state-of-the-art models formalize KGC predictions as functions of not just the target triple, but also of rich local and global context.

Entity neighborhood context: For a query $(h, r, ?)$ , one-hop neighbors of $h$ are encoded as $(r_i, e_i)$ pairs (Chen et al., 29 Mar 2025).
Relation context: Triples sharing relation $r$ , $(h_j, r, t_j)$ , are included to capture occurrence and diversity patterns of $r$ across the KG.
Input fusion: Contextual elements are concatenated as tokenized text, segmented into blocks, and fed to transformer encoders or bi-encoder models. Models such as KGC-ERC (Chen et al., 29 Mar 2025) and MuCo-KGC (Gul et al., 5 Mar 2025) employ careful sampling strategies—grouping, clustering, and prioritization—underfixed token budgets (typically 512 tokens).

Mathematical formalization:

Let $G$ be the KG with entity set $E$ , relation set $R$ .

Entity neighborhood: $\text{En}(q) = \{ (r_i, e_i) \mid (h, r_i, e_i) \in G \}$ .
Relation context: $\text{Rc}(q) = \{ (h_j, r, t_j) \mid (h_j, r, t_j) \in G \}$ .

Final input $q'$ for a generative LM is:

$q' = \text{concat}(\text{query: } h | r, \text{SEP}, \text{entity neighborhood:}, \text{En}(q), \text{SEP}, \text{relation context:}, \text{Rc}(q))$

Tokenization employs SentencePiece or BERT-style encoders (Chen et al., 29 Mar 2025, Gul et al., 2024).

3. Training Objectives, Inference, and Scalability

Training: Recent models eliminate negative sampling, leveraging cross-entropy over all candidate entities for tail (or head) prediction. This is both more efficient and avoids bias from false negatives (Gul et al., 2024, Gul et al., 5 Mar 2025). For a target sequence $Y = (y_1, \ldots, y_T)$ , the objective is:

$L = - \sum_{t=1}^T \log p(y_t | y_{<t}, q')$

Inference: Instead of exhaustively scoring all $|E|$ candidates, generative decoders sample a fixed subset of output sequences and map text back to entities, achieving sub-linear runtime (Chen et al., 29 Mar 2025). Bi-encoder architectures precompute entity embeddings for fast lookup; transformer-based models utilize attention for context spreading across segments.

Sampling strategies: Customized selection of context triples—by cardinality type, clustering, or priority—ensures maximally informative yet compact context (Chen et al., 29 Mar 2025). Ablations on context selection indicate nontrivial MRR drops when relation context or sampling rules are removed.

4. Empirical Results and Performance Analysis

Robust experimental protocols employ large KGs (Wikidata5M (Chen et al., 29 Mar 2025), FB15K-237-N, Wiki27K, WN18RR (Gul et al., 2024, Gul et al., 5 Mar 2025)) and standard metrics:

Mean Reciprocal Rank (MRR)
Hits@K (K = 1, 3, 10)
Filtered setting (excluding other true triples)

Performance highlights:

On Wikidata5M, KGC-ERC achieves MRR 0.386, Hits@1 0.360, best among text-only methods, and approaches structure-based SOTA (Chen et al., 29 Mar 2025).
Appending entity descriptions further increases MRR to 0.433, Hits@1 0.412 for KGC-ERC+Desc.
On FB15K-237-N and Wiki27K, KGC-ERC surpasses strong baselines (TuckER, PKGC, VKGC).
CAB-KGC (Gul et al., 2024) attains new state-of-the-art (MRR 0.350 on FB15k-237, MRR 0.685 on WN18RR), with significant boosts (+5.3% and +4.88% in Hits@1).
MuCo-KGC (Gul et al., 5 Mar 2025) outperforms all prior methods on three benchmarks (i.e. WN18RR, CoDEx-S, CoDEx-M), with up to +20.15% improvement in MRR.

5. Scalability, Complexity, and Ablation Studies

Text-based and generative LM methods provide major advantages in scalability:

Inference complexity: Generative models such as KGC-ERC decode only a sample of candidate entities, yielding inference time that is sub-linear in $|E|$ and competitive with bi-encoder methods for $|E| \gtrsim 10^5$ (Chen et al., 29 Mar 2025).
Training overhead: Models eschewing negative sampling (MuCo-KGC, CAB-KGC) reduce computational demands and data imbalance issues (Gul et al., 2024, Gul et al., 5 Mar 2025).
Parameter efficiency: MuCo-KGC (110M parameters) matches or exceeds much larger LLM-based systems (up to 7B parameters).
Ablations: Removing relation context (Rc(q)) or tailored sampling rules in KGC-ERC results in measurable drops in predictive accuracy (–1.1 MRR and –0.4 MRR, respectively) (Chen et al., 29 Mar 2025).

6. Analysis of Context Utilization and Model Capabilities

Empirical evidence and ablation studies indicate context fusion in input encoding is critical for disambiguation, prediction accuracy, and generalization.

Integrated entity neighborhood and relation context provide complementary signals. Relation context introduces disambiguating factors not present in local neighborhoods (Chen et al., 29 Mar 2025).
Balanced fusion—via BERT or T5 self-attention—enables the model to attend effectively across “query,” entity context, and relation context blocks.
Models dispensing with textual descriptions (MuCo-KGC, CAB-KGC) demonstrate robust performance on text-poor graphs (CoDEx), highlighting the sufficiency of structured contextual signals (Gul et al., 5 Mar 2025).
Growing evidence favors context-aware architectures and sampling over mere scaling of model/parameter size for performance increase.

7. Future Directions and Open Questions

Several lines of inquiry remain active:

Dynamic multi-hop neighborhood sampling and path-based context construction may further boost performance on complex KGs (Gul et al., 5 Mar 2025).
Lightweight transformers or knowledge-dense pre-training for improved scalability versus BERT-based architectures.
Adaptive or learned context fusion mechanisms remain largely unexplored.
Extending context-aware KGC systems to multilingual, cross-modal, or temporal KGs is critical for broad applicability.

A plausible implication is that further improvements will primarily arise from more sophisticated exploitation of graph structure and context, rather than from mere scaling of LLMs or KG sizes.

Principal references: Enhancing Knowledge Graph Completion with Entity Neighborhood and Relation Context (Chen et al., 29 Mar 2025), A Contextualized BERT model for Knowledge Graph Completion (Gul et al., 2024), MuCo-KGC: Multi-Context-Aware Knowledge Graph Completion (Gul et al., 5 Mar 2025).