Abstract
Graph is a ubiquitous structure in natural language processing (NLP), which describes a collection of entities, represented as nodes, and their pairwise relationships, represented as edges. Many sentence-level meaning representations employ directed, acyclic graphs as the underlying formalism, while most tree-based syntactic representations can also be regarded as graphs. In this thesis, we mainly focus on predicting graphs from texts, and integrating graphs for downstream tasks, such as natural language generation and relation extraction. We first investigate how to predict the corresponding semantic graph from the input sentence. Specifically, we focus on the Abstract Meaning Representation (AMR), which is a directed, labelled graph. Such a graph can capture rich structural information, and is useful for semantics-related tasks. AMR parsing is a challenging task as it requires the parser to learn to predict a large number of concepts and relations based on relatively sparse training data. We introduce a simple yet effective transition-based AMR parser, which is able to conduct the search in a refined search space based on a new compact AMR graph and an improved oracle. Then we explore how to leverage graph structures for natural language generation. The key challenge in graph encoding is how to efficiently learn an effective graph representation. We introduce dense connections and the dynamic fusion mechanism to graph convolutional networks (GCNs). With the help of these mechanisms, the GCN model is able to capture rich local and non-local information associated with a graph. Two novel weight sharing strategies are further developed to reduce memory usage and model complexity. Apart from language generation, relation extraction models also benefit from incorporating graph structural information. A key challenge in graph-based models is the error propagation. To better address this issue, we propose a soft pruning strategy that is able to distil the information associated with the dependency graph. For document-level and bio-medical texts, the error propagation issue becomes severe due to the low parsing accuracy. We further develop models based on latent structure induction, which treats the dependency graph as a latent variable and induces it in an end-to-end fashion.