Triple-based graph neural network for encoding event units in graph reasoning problems

Abstract Nowadays graph neural networks have achieved excellent performance on many graph-based tasks such as abstract meaning representation (AMR) text generation and graph reasoning. Graph-based models often calculate the information flow by nodes and their associated edges. But node-based or edge-based calculation can not reflect the strong relation between every two nodes as triples, which actually act as event units in graph reasoning tasks. Considering triples (a triple means an edge and its two nodes in our model) as basic calculation units is more suitable for relation-based graphs and event-reasoning tasks. We thus propose a novel structure called triple-based graph neural network and directly perform triple-based neural calculation on interrelated triples. Experimental results on babi 16, babi 19, wikisql and our own proposed genealogy dataset show that our model significantly outperforms other strong alternatives.

Paper

Full text

PDF

Triple-based graph neural network for encoding event units in graph reasoning problems

Semantic Scholar · Computer Science · 2021

Abstract

Abstract Nowadays graph neural networks have achieved excellent performance on many graph-based tasks such as abstract meaning representation (AMR) text generation and graph reasoning. Graph-based models often calculate the information flow by nodes and their associated edges. But node-based or edge-based calculation can not reflect the strong relation between every two nodes as triples, which actually act as event units in graph reasoning tasks. Considering triples (a triple means an edge and its two nodes in our model) as basic calculation units is more suitable for relation-based graphs and event-reasoning tasks. We thus propose a novel structure called triple-based graph neural network and directly perform triple-based neural calculation on interrelated triples. Experimental results on babi 16, babi 19, wikisql and our own proposed genealogy dataset show that our model significantly outperforms other strong alternatives.

Similar papers

© 2026 NYSGPT2525 LLC