Relational Context Enhanced Dual Path Reasoning for Fact Prediction in Knowledge Graph.
Yilin Wang,Zhen Huang,Minghao Hu,Dongsheng Li,Wei Luo,Dong Yang
DOI: https://doi.org/10.1109/smartworld-uic-atc-scalcom-digitaltwin-pricomp-metaverse56740.2022.00207
2022-01-01
Abstract:Path-based knowledge graph reasoning approaches, which typically encode paths from the head entity to the tail entity for fact prediction, show good performance. Previous work demonstrates that leveraging entity type information can bring in benefits for learning path representations. However, many entities are not typed or loosely typed, resulting in inaccurate understanding of entity semantics. In addition, prior methods tend to model paths only from the forward direction through sequential networks like LSTM. Thus, they can neither observe the path information after the current encoding step nor capture new path patterns from the reverse direction. To address the above problems, we introduce a Relational Context enhanced Dual Path Reasoning (RCDPR) model in this paper. First, the model utilizes relational contexts, namely the relations linked from the given entity, to derive reliable entity semantics. Second, Transformer-based path encoder is designed to obtain path representations, so as to look into the information of the whole path and apply self-attention to emphasize important entity semantics. A dual path encoding method is presented to enhance the path representation by additionally encoding the new patterns embedded in the reverse path, which are combined with the forward path representation for prediction. Experimental results on two benchmark datasets, WN18RR and FB15k-237, show that our approach outperforms state-of-the-art methods on the fact prediction task, which obtains absolute MAP improvements of 2.62% and 3.66%, respectively, compared to path-based methods. Furthermore, adequate experiments illustrate the effectiveness of utilizing entity relational contexts and dual path encoding for inference.