Core Concepts
A relation prediction model that integrates structural and textual embeddings to effectively complete knowledge graphs.
Abstract
The paper proposes a relation prediction model (RPEST) that utilizes both the structural information and the textual content of knowledge graph nodes to enhance knowledge graph completion.
The key highlights are:
The model employs a walk-based graph structure algorithm (Node2Vec) to generate structural embeddings, replacing the costly fine-tuning step in masked language models.
It exploits pre-trained language models (Glove) to capture text contextualized representation, avoiding the high computational overhead of fine-tuning large language models.
The model integrates the structural and textual embeddings through a neural network architecture that includes a bidirectional LSTM layer, an attention layer, and a prediction layer.
Experiments on the FB15K dataset show that the proposed RPEST model achieves competitive results compared to state-of-the-art relation prediction models, outperforming them on several evaluation metrics.
An ablation study demonstrates the effectiveness of incorporating both structural and textual information, with the Glove-based variant performing better than the BERT-based one in terms of both performance and efficiency.
Overall, the paper presents a novel approach that effectively leverages the complementary strengths of structural and textual information to enhance knowledge graph completion, particularly in the relation prediction task.
Stats
The average entity name length in the Freebase dataset is 2.7 words.
Less than 2% of the words in the datasets are out-of-vocabulary when using Glove.
Quotes
"Our model computes the neural network training loss using the cross entropy loss function."
"We reason our model's superiority by the combination of the structural and textual details for every node."