Entity disambiguation method based on Graph Attention Networks

Entity disambiguation based on entity-link is a technique which constructs the mappings between entity reference items appearing in the short text and target entity in knowledge base respectively. In this paper we propose an entity disambiguation method based on Graph Attention Networks for semi-structured knowledge base data. First, a global Knowledge Graph is constructed from the semi-structured knowledge base, and the entity reference items are embedded by Bert pre-trained model meanwhile. Next, Graph Attention Networks which leverages masked self-attention layers is applyed on candidate entity nodes of global Knowledge Graph to fetch a vector of node level. Furtherly, we compute similarity scores rank between the entity reference items and the candidate entity to complete the task of entity disambiguation. The experimental results on CCKS2019 dataset achieve state-of-the-art.

I. Introduction

Entity disambiguation based on entity-link can resolve the problem of multiple meanings of entities [1]. As illustrated in Figure 1, the entity "匕里香"(“Seven Miles of Fragrance”) has multiple implication in the knowledge base. In recent years, many methods of entity disambiguation are based on text [2], but these don’t consider the relationship between sequences. Consequently, Knowledge Graph mainly utilizes the graphical data structure to represent node’s relationships [3], and provides a better solution for entity disambiguation. Figure 2 gives an example, According to the contextual features of the Knowledge Graph nodes, we can link text’s “Seven Miles of Fragrance” to “Seven Miles of Fragrance” being in the Knowledge Graph that aims at eliminating the ambiguity caused by other meaning items.