Progressive Knowledge Graph Completion

TL;DR

This work reframes Knowledge Graph Completion as Progressive Knowledge Graph Completion (PKGC), integrating training, mining, and verification to mirror real-world KG construction constraints and verifier limits. It introduces two acceleration modules, Optimized Top-$k$ and Semantic Validity Filter, to dramatically speed up candidate mining while maintaining effectiveness. Through experiments on FB15k and WN18-rescal-based setups, the authors show that traditional link-prediction metrics do not predict PKGC performance, with UniBi variants—especially when normalization is applied—achieving strong results. The study also explores incremental learning and low-resource scenarios, demonstrating PKGC's practical viability and identifying directions for future improvements in realism-aware KG completion. PKGC thus offers a more realistic, dynamic framework for evaluating and developing KGC methods in real-world settings.

Abstract

Knowledge Graph Completion (KGC) has emerged as a promising solution to address the issue of incompleteness within Knowledge Graphs (KGs). Traditional KGC research primarily centers on triple classification and link prediction. Nevertheless, we contend that these tasks do not align well with real-world scenarios and merely serve as surrogate benchmarks. In this paper, we investigate three crucial processes relevant to real-world construction scenarios: (a) the verification process, which arises from the necessity and limitations of human verifiers; (b) the mining process, which identifies the most promising candidates for verification; and (c) the training process, which harnesses verified data for subsequent utilization; in order to achieve a transition toward more realistic challenges. By integrating these three processes, we introduce the Progressive Knowledge Graph Completion (PKGC) task, which simulates the gradual completion of KGs in real-world scenarios. Furthermore, to expedite PKGC processing, we propose two acceleration modules: Optimized Top-$k$ algorithm and Semantic Validity Filter. These modules significantly enhance the efficiency of the mining procedure. Our experiments demonstrate that performance in link prediction does not accurately reflect performance in PKGC. A more in-depth analysis reveals the key factors influencing the results and provides potential directions for future research.

Figures (13)

• Figure 1: PKGC consists of training, mining and verification procedures. The knowledge proposed by the KGE model will be added to the knowledge base after verification.
• Figure 2: Figure (a) depicts the Root filter process, where lower-scoring triplets are directly filtered out. On the right, Figure (b) demonstrates Batch warm-up. In this process, the data from the initial batch undergoes decomposition, and its size gradually increases until it reaches a predetermined limit. And subsequent batches maintain a consistent size.
• Figure 3: Illustration of how to calculate MOAR, which is the ratio of the area enclosed by the actual curve and the ideal curve, respectively.
• Figure 4: Figures illustrating the dynamic completion process for various models on the WN18 and FB15k datasets. The closer the trend of the curve is to the upper left, the more efficient the model is in performing the dynamic completion. It is evident that UniBi-O(2) and UniBi-O(3) maintain a significant advantage on both datasets, while TransE performs poorly on both.
• Figure 5: Ablation studies of relation normalization (RelNorm) on (a) WN18 and (b) FB15k. We utilize CP and ComplEx as examples.