SARF: Aliasing Relation Assisted Self-Supervised Learning for Few-shot Relation Reasoning

TL;DR

A novel Self-supervised learning model by leveraging Aliasing Relations to assist FS-KGR, termed SARF is proposed, and extensive experiments demonstrate that SARF achieves state-of-the-art (SOTA) performance compared with other methods in most cases.

Abstract

Few-shot relation reasoning on knowledge graphs (FS-KGR) aims to infer long-tail data-poor relations, which has drawn increasing attention these years due to its practicalities. The pre-training of previous methods needs to manually construct the meta-relation set, leading to numerous labor costs. Self-supervised learning (SSL) is treated as a solution to tackle the issue, but still at an early stage for FS-KGR task. Moreover, most of the existing methods ignore leveraging the beneficial information from aliasing relations (AR), i.e., data-rich relations with similar contextual semantics to the target data-poor relation. Therefore, we proposed a novel Self-Supervised Learning model by leveraging Aliasing Relations to assist FS-KGR, termed SARF. Concretely, four main components are designed in our model, i.e., SSL reasoning module, AR-assisted mechanism, fusion module, and scoring function. We first generate the representation of the co-occurrence patterns in a generative manner. Meanwhile, the representations of aliasing relations are learned to enhance reasoning in the AR-assist mechanism. Besides, multiple strategies, i.e., simple summation and learnable fusion, are offered for representation fusion. Finally, the generated representation is used for scoring. Extensive experiments on three few-shot benchmarks demonstrate that SARF achieves state-of-the-art performance compared with other methods in most cases.

Figures (7)

• Figure 1: The statistics of relation frequencies of FB15K-237. FB15K-237 contains 237 relations and 310116 triplets. There are a large amount of long-tail relations with only few instances.
• Figure 2: Illustration of aliasing relations. Aliasing relations ("ParentOf" and "MotherOf" in right figure) are selected from background KG over semantically similarity comparison with the target relation ("FatherOf" in left figure).
• Figure 3: The framework of SARF. The SARF includes four main steps: self-supervised reasoning module, aliasing relation assisted module, fusion module, and triplet scoring module. Concretely, in the first step, we construct the few-shot tasks including a support set and a query set corresponding to the target relation (orange dotted line). In the second step, the co-occurrence patterns are extracted over a co-occurrence extractor $\rm{C_E}$ to represent the target relation. After that, the co-occurrence patterns will be encoded over a graph encoder; In the third step, we select aliasing relations corresponding to the target relation to extract (AR-subgraphs) and encode them into representations. Finally, we fuse the representation of the co-occurrence pattern and AR-subgraphs and test whether there is evidence close enough to the co-occurrence pattern by using a reconstructor $\rm{C_R}$.In general, all the edges shown represent different relations. But in this figure, we specifically highlight the aliasing relations with colors.
• Figure 4: The comparison of data preparing process between meta-Learning based FS-KGR methods and SSL-based FS-KGR methods on FB15K-237. The meta-learning based FS-KGR method need to design 207 training tasks, 30 support sets, and 30 query sets manually before training. While SSL-based FS-KGR methods only need 30 support sets and 30 query sets.
• Figure 5: Computational cost analysis of our model and two other baselines in training process. Concretely, We compare the inference time (seconds) of 1 training epoch on three datasets, including NELL, FB15K-237 and ConceptNet.