Two papers were recently accepted to the ACL conference, showcasing innovative approaches to address the challenge of predicting missing links in incomplete knowledge graphs.
Knowledge graphs are a collection of facts represented as triples. For instance, the fact that Berlin is the capital of Germany can be expressed with the triple (Berlin, capital of, Germany), where "Berlin "and “Germany" represent the head and tail entities, respectively, and "capital of" is a relation name. One of the challenges in knowledge graph-based applications is dealing with incomplete data. A prominent approach for predicting missing links are knowledge graph embeddings, where entities and relation names are mapped to a vector space.
Nayyeri et al. introduce a novel method for knowledge graph embeddings to capture intricate graph patterns and leverage them for predicting missing links. In particular, knowledge graphs encompass a blend of branching relations, e.g., the relation “friend of”, where a person can be friend of many other people, and complex structural patterns, e.g., “part of”, which represents an example of a tree-like relation. Current embedding methods employ a pointwise transformation to map graph nodes onto a vector space and encounter difficulties in handling these kinds of patterns. Nayyeri et al. address this concern by modeling relations between nodes using relation-specific stochastic transitions. The proposed method, called ItoE, facilitates the modeling of diverse relationships in knowledge graphs, including branching relations and structural patterns.
In Xiong et al., the focus was on hyper-relational knowledge graphs. Hyper-relational knowledge graphs incorporate primal triple and additional qualifiers that provide contextual information. For example, the primal triple (Einstein, educated at, Zurich) can have qualifiers specifying that his major was Physics and he obtained a PhD degree. The authors proposed a novel method, called ShrinkE that surpasses existing state-of-the-art approaches in link prediction for hyper-relational graphs. This method is able to model both triple-level patterns, such as symmetry, and qualifier-level patterns, such as qualifier implication. By considering those patterns, the proposed technique achieves superior performance in predicting links within hyper-relational knowledge graphs.
M. Nayyeri, B. Xiong, M. Mohammadi, M. M. Akter, M. M. Alam, J. Lehmann, S. Staab. Knowledge Graph Embeddings using Neural Itô Process: From Multiple Walks to Stochastic Trajectories. In: Findings of the Association for Computational Linguistics: ACL 2023, Toronto, Canada, July 9-14, 2023.
B. Xiong, M. Nayyeri, S. Pan, S. Staab. Shrinking Embeddings for Hyper-Relational Knowledge Graphs. In: Proc. Of The 61st Annual Meeting of the Association for Computational Linguistics (ACL 2023), Toronto, Canada, July 9-14, 2023.