OpticE: A Coherence Theory-Based Model for Link Prediction

Proceedings of the 29th International Conference on Computational Linguistics, pages 4892–4901 October 12–17, 2022. 4892 OpticE: A Coherence Theory-Based Model for Link Prediction Xiangyu Gui, Feng Zhao ∗, Langjunqing Jin, Hai Jin National Engineering Research Center for Big Data Technology and System, Services Computing Technology and System Lab, Cluster and Grid Computing Lab School of Computer Science and Technology, Huazhong University of Science and Technology, China zhaof@hust.edu.cn Abstract Knowledge representation learning is a key step required for link prediction tasks with knowledge graphs (KGs). During the learn- ing process, the semantics of each entity are embedded by a vector or a point in a feature space. The distance between these points is a measure of semantic similarity. However, in a KG, while two entities may have simi- lar semantics in some relations, they have dif- ferent semantics in others. It is ambiguous to assign a fixed distance to depict the vari- ant semantic similarity of entities. To allevi- ate the semantic ambiguity in KGs, we design a new embedding approach named OpticE, which is derived from the well-known physi- cal phenomenon of optical interference. It is a lightweight and relation-adaptive model based on coherence theory, in which each entity’s semantics vary automatically regarding differ- ent relations. In addition, a unique negative sampling method is proposed to combine the multimapping properties and self-adversarial learning during the training process. The ex- perimental results obtained on practical KG benchmarks show that the OpticE model, with elegant structures, can compete with existing link prediction methods. 1 Introduction Knowledge graphs (KGs) consist of sets of triplets that can represent real-world concepts, common sense information or facts. Each triplet (h, r, t) in- dicates a directional relation r from the head en- tity h to the tail entity t. WordNet (Miller, 1995), Freebase (Bollacker et al., 2008) and Wikidata (Vrandeˇci´c and Krötzsch, 2014) are the most well- known KGs with tremendous numbers of entities and relations. These KGs play important roles in a range of areas, such as natural language process- ing tasks (Cao et al., 2021), question answering ap- plications (Bosselut et al., 2021) and recommenda- tion systems (Shao et al., 2021). Research on KGs ∗Corresponding author is thriving in both academic and industrial commu- nities. Link prediction is a fundamental problem when addressing KGs. To learn the hidden patterns from the observed triplets, extensive investigations have been performed to embed entities and re- lations into a continuous semantic space; this is known as KG embedding learning (Rossi et al., 2021). Among these embedding models, from a mathematical perspective, TransE (Bordes et al., 2013), TransH (Wang et al., 2014) and TransR (Lin et al., 2015) are translation-based models that use translation transformations to character- ize the existence of triplets. Other methods (e.g., RESCAL (Nickel et al., 2011) and DistMult (Yang et al., 2015)) tackle this problem with matrix or tensor multiplication. Recently, the embedding problem has been discussed in the complex do- main (e.g., ComplEx (Trouillon et al., 2016), Ro- tatE (Sun et al., 2019) …

Content truncated to 500 words