Deep learning of knowledge graph embeddings for semantic parsing of Twitter dialogs

Abstract

This paper presents a novel method to learn neural knowledge graph embeddings. The embeddings are used to compute semantic relatedness in a coherence-based semantic parser. The approach learns embeddings directly from structured knowledge representations. A deep neural network approach known as Deep Structured Semantic Modeling (DSSM) is used to scale the approach to learn neural embeddings for all of the concepts (pages) of Wikipedia. Experiments on Twitter dialogs show a 23.6% reduction in semantic parsing errors compared to the state-of-the-art unsupervised approach.