Meta-Semantics Augmented Few-Shot Relational Learning

TL;DR

PromptMeta addresses the challenge of few-shot relational learning on knowledge graphs by integrating meta-semantic knowledge with relational structure. It introduces a Meta-Semantic Prompt pool and a learnable fusion mechanism to transfer semantic cues across tasks and adapt to unseen relations in meta-learning. The approach combines context-aware neighbor aggregation, semantic prompts, and a fusion token to form task-specific meta-representations, optimized with a joint objective that includes pool tuning. Empirical results on Nell-One and Wiki-One show consistent improvements over state-of-the-art baselines, with ablations confirming the value of each component and semantic embeddings for further gains.

Abstract

Few-shot relational learning on knowledge graph (KGs) aims to perform reasoning over relations with only a few training examples. While current methods have focused primarily on leveraging specific relational information, rich semantics inherent in KGs have been largely overlooked. To bridge this gap, we propose PromptMeta, a novel prompted meta-learning framework that seamlessly integrates meta-semantics with relational information for few-shot relational learning. PromptMeta introduces two core innovations: (1) a Meta-Semantic Prompt (MSP) pool that learns and consolidates high-level meta-semantics shared across tasks, enabling effective knowledge transfer and adaptation to newly emerging relations; and (2) a learnable fusion mechanism that dynamically combines meta-semantics with task-specific relational information tailored to different few-shot tasks. Both components are optimized jointly with model parameters within a meta-learning framework. Extensive experiments and analyses on two real-world KG benchmarks validate the effectiveness of PromptMeta in adapting to new relations with limited supervision.

Figures (2)

• Figure 1: Illustration of PromptMeta's meta-training process. Entities are initialized with pre-trained relational embeddings and enriched via context-aware neighbor aggregation to capture local relational information. The self-attention function $f_\textrm{sa}$ computes task-relational embedding $\mathbf{r}^{(r)}$, while task-semantic prompt $\mathbf{r}^{(p)}$ is retrieved from the MSP pool. These are fused through $\Phi_\textrm{fuse}$ with the fusion token $\mathbf{r}^{(f)}$ to generate the meta-representation $\mathbf{r}^m$. The support loss $\mathcal{L}(S_r)$ is computed to optimize model parameters, after which the updated meta-representation $\mathbf{r}^{m'}$ is adapted to the query set by optimizing the overall training loss $\mathcal{L}$.
• Figure 2: t-SNE visualization and case study.