Soft Reasoning Paths for Knowledge Graph Completion

TL;DR

This paper tackles the problem of knowledge graph completion when reasoning paths are incomplete or unavailable. It introduces soft reasoning paths by attaching trainable embeddings to each relation and aligns these embeddings with actual reasoning paths through a contrastive objective, complemented by a dual-encoder architecture. A hierarchical ranking strategy combines signals from relations, reasoning paths, and soft paths to improve both efficiency and accuracy, demonstrated across WN18RR, FB15k-237, and Wikidata5M-Trans with state-of-the-art results and robustness to missing-path scenarios. While the approach incurs higher computational costs, the substantial performance gains suggest a favorable trade-off for large-scale KGC applications, with future work aimed at further efficiency improvements.

Abstract

Reasoning paths are reliable information in knowledge graph completion (KGC) in which algorithms can find strong clues of the actual relation between entities. However, in real-world applications, it is difficult to guarantee that computationally affordable paths exist toward all candidate entities. According to our observation, the prediction accuracy drops significantly when paths are absent. To make the proposed algorithm more stable against the missing path circumstances, we introduce soft reasoning paths. Concretely, a specific learnable latent path embedding is concatenated to each relation to help better model the characteristics of the corresponding paths. The combination of the relation and the corresponding learnable embedding is termed a soft path in our paper. By aligning the soft paths with the reasoning paths, a learnable embedding is guided to learn a generalized path representation of the corresponding relation. In addition, we introduce a hierarchical ranking strategy to make full use of information about the entity, relation, path, and soft path to help improve both the efficiency and accuracy of the model. Extensive experimental results illustrate that our algorithm outperforms the compared state-of-the-art algorithms by a notable margin. The code will be made publicly available after the paper is officially accepted.

Figures (2)

• Figure 1: SRP-KGC Framework: During the training process, we introduced three types of positive samples. By incorporating these diverse positive samples, the model's ability to understand reasoning paths was enhanced, while the soft reasoning path learns the generalized representation of reasoning paths. In the testing phase, we employed a hierarchical ranking strategy, combining information from entities, relations, soft reasoning paths, and reasoning paths to further improve the model's accuracy.
• Figure 2: Visualization of embeddings with the different head entities and relations using t-SNE under the settings of SimKGC and SRP-KGC. In the visualization, points with the same color represent embeddings that share the same target tail entity.