Context-Aware Adapter Tuning for Few-Shot Relation Learning in Knowledge Graphs

Ran Liu,Zhongzhou Liu,Xiaoli Li,Yuan Fang

2024-10-17

Abstract:Knowledge graphs (KGs) are instrumental in various real-world applications, yet they often suffer from incompleteness due to missing relations. To predict instances for novel relations with limited training examples, few-shot relation learning approaches have emerged, utilizing techniques such as meta-learning. However, the assumption is that novel relations in meta-testing and base relations in meta-training are independently and identically distributed, which may not hold in practice. To address the limitation, we propose RelAdapter, a context-aware adapter for few-shot relation learning in KGs designed to enhance the adaptation process in meta-learning. First, RelAdapter is equipped with a lightweight adapter module that facilitates relation-specific, tunable adaptation of meta-knowledge in a parameter-efficient manner. Second, RelAdapter is enriched with contextual information about the target relation, enabling enhanced adaptation to each distinct relation. Extensive experiments on three benchmark KGs validate the superiority of RelAdapter over state-of-the-art methods.

Machine Learning,Artificial Intelligence

What problem does this paper attempt to address?

The problem that this paper attempts to solve is how to deal with the distribution differences between new relations and the relations seen during the training process when conducting few - shot relation learning (FSRL) in knowledge graphs (KGs). Specifically:

1. **Limitations of existing methods**:
- Most of the existing FSRL methods are based on meta - learning and assume that the tasks in meta - training and meta - testing are independent and identically distributed (i.i.d.). However, in practical applications, the distributions of different relations may vary greatly, which makes this assumption no longer valid.
- Such distribution differences will lead to a decline in the performance of the model when dealing with unseen new relations, especially in the few - shot case.

2. **Problems proposed in the paper**:
- How to design a model that can adapt to different relation distributions, especially in the meta - testing stage, and can quickly adapt to new relations with a small number of samples.
- How to use context information to enhance the learning of new relations, thereby improving the generalization ability of the model.

To solve these problems, the paper proposes RelAdapter, a framework based on context - aware adapters, aiming to improve FSRL in the following ways:

- **Lightweight adapter module**: By introducing a lightweight adapter module, specific - relation adjustment of global prior knowledge is achieved while maintaining parameter efficiency.
- **Context information enhancement**: Utilize the existing structural information in the knowledge graph to extract the context information of the target relation and further enhance the adaptability to new relations.

These improvements make RelAdapter significantly outperform the existing FSRL methods on multiple benchmark datasets.