核心概念
This paper introduces a new task called Few-Shot Link Prediction on Hyper-relational Facts (FSLPHFs), which aims to predict a missing entity in a hyper-relational fact with limited support instances. The authors propose MetaRH, a model that learns Meta Relational information in Hyper-relational facts to accurately predict the missing entity.
要約
The paper introduces the task of Few-Shot Link Prediction on Hyper-relational Facts (FSLPHFs), which aims to predict a missing entity in a hyper-relational fact with limited support instances. To tackle this task, the authors propose MetaRH, a model with three key modules:
-
Relation Learning Module:
- Generates initial few-shot relation representations by aggregating entity background facts and encoding support instances.
- Utilizes a Graph Neural Network with attention and gating mechanisms to enhance entity representations using background facts.
- Employs the GRAN model to generate the few-shot relation representation.
-
Support-specific Adjustment Module:
- Adjusts the coarse relation representation based on the support set to obtain meta relational information.
- Introduces an instance scorer to evaluate the semantic connections between few-shot relations and other elements in instances.
- Utilizes the gradient on support instances to guide the adjustment of the relation representation.
-
Query Inference Module:
- Predicts the missing entity in a query using the obtained meta relational information.
- Adapts the same instance scorer structure and shares parameters as the one in the support-specific adjustment module.
The authors construct three new datasets, F-WikiPeople, F-JF17K, and F-WD50K, based on existing LPHFs benchmark datasets to evaluate the effectiveness of MetaRH. Experimental results demonstrate that MetaRH significantly outperforms existing representative models on these datasets.
統計
32.5% of relations in the WD50K dataset have less than 5 instances.
In the F-JF17K dataset, 49.3% of the facts are hyper-relational.
In the F-WD50K dataset, 13.8% of the facts are hyper-relational.