Text-Graph Enhanced Knowledge Graph Representation Learning.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Data Science

Background:

• Knowledge Graphs (KGs) are crucial for NLP tasks, but traditional embedding methods struggle with sparse data.
• Existing methods incorporating entity descriptions often overlook global word co-occurrence.
• There is a need for KG embedding techniques that leverage both local and global textual context.

Purpose of the Study:

• To develop an end-to-end KG embedding model that addresses structure sparsity by incorporating auxiliary text information.
• To enhance KG representations by modeling auxiliary texts as a graph capturing both local and global semantic relationships.
• To improve the performance of KG embedding methods on benchmark datasets.

Main Methods:

• Proposed Teger, an end-to-end text-graph enhanced KG embedding model.
• Modeled auxiliary entity texts using a heterogeneous entity-word graph (text-graph) to capture local and global semantics.
• Applied graph convolutional networks (GCNs) to learn entity embeddings from the text-graph.
• Integrated text-graph embeddings with KG triplet embeddings using a gating mechanism.

Main Results:

• The Teger model significantly outperforms several state-of-the-art KG embedding methods.
• Incorporating global word co-occurrence information from auxiliary texts proved beneficial.
• The proposed text-graph approach effectively enriches KG representations and alleviates structure sparsity.

Conclusions:

• Teger offers a robust solution for KG embedding by effectively utilizing auxiliary text data.
• The method demonstrates the importance of modeling both local and global textual semantics for improved KG representation learning.
• This approach enhances KG quality and performance in various downstream NLP applications.