Graph Deep Learning for Natural Language Processing

Language, by its very nature, is inherently structured and relational. Words form sentences, and sentences form documents, which contain concepts that interlink in complex ways to convey meaning. Graph structures provide an ideal framework to capture these intricate relationships, going beyond the traditional models. By representing text as graphs, we can leverage the rich expressiveness of graph theory and the computational power of deep learning to tackle challenging natural language processing (NLP) problems.

In this chapter, we will delve into the fundamental concepts of graph representations in NLP, exploring various types of linguistic graphs such as dependency trees, co-occurrence networks, and knowledge graphs. We’ll then build upon this foundation to examine the architectures and mechanisms of graph neural networks (GNNs) that have been specifically adapted for language tasks.

We’ll cover the following...