Regular term determination method and device for tensor decomposition type knowledge graph completion

Abstract

The embodiment of the invention discloses a regular term determination method and device for tensor decomposition type knowledge graph completion. The method comprises the following steps: determining an original model, wherein the original model isa scoring function based on a tensor decomposition model; determining a dual model corresponding to the original model, wherein the dual model is a model scoring function based on distance; determining duality between the original model and the dual model; and determining a first regular term corresponding to the original model based on the duality. After the first regular term is determined, the first regular term can be subsequently applied to tensor decomposition-oriented knowledge graph completion, so that the overfitting problem is prevented while the expressivity of a tensor decomposition-based knowledge graph completion model is kept; and meanwhile, the method can be widely applied to various models based on tensor decomposition, and has wider application range.

Description

technical field [0001] The present invention relates to natural language processing technology, and more specifically, relates to a method and device for determining regularization items oriented to the completion of tensor decomposition-type knowledge graphs. Background technique [0002] Knowledge graphs consist of a large number of fact triples, which store structured human knowledge. In recent years, knowledge graphs have made great achievements in fields such as natural language processing, intelligent question answering, recommender systems, and computer vision. [0003] Commonly used knowledge graphs usually contain billions of triples, but there are still a large number of missing triples. Due to the extremely large scale of the knowledge graph, it is expensive to manually complete the graph. Therefore, the knowledge graph completion technology that automatically predicts missing triples based on known triples in the knowledge graph has attracted much attention in ...

AI Technical Summary

Problems solved by technology

Distance-based models use Minkowski distance to measure the rationality of triplets. Although such models can achieve the best performance so far, they are still difficult to model when modeling complex relational patterns (such as one-to-many and many-to-one relationships). There are difficulties

Tensor decomposition-based models treat knowledge graphs as partially observable third-order tensors, thereby modeling knowledge graph completion as a tensor completion problem; theoretically, these models are expressive and can perform well handle complex relationships efficiently, however they often suffer from serious overfitting problems and cannot achieve optimal performance

[0005] In order to solve the overfitting problem based on the tensor decomposition model, researchers have proposed various regularization methods, but there is currently no ideal regularization method that is both efficient and widely applicable.

Images