Graph-based Approaches to Text Generation

Deep Learning advances have enabled more fluent and flexible text generation. However, while these neural generative approaches were initially successful in tasks such as machine translation, they face problems – such as unfaithfulness to the source, repetition and incoherence – when applied to generation tasks where the input is structured data, such as graphs. Generating text from graph-based data, including Abstract Meaning Representation (AMR) or Knowledge Graphs (KG), is a challenging task due to the inherent difficulty of properly encoding the input graph while maintaining its original semantic structure. Previous work requires linearizing the input graph, which makes it complicated to properly capture the graph structure since the linearized representation weakens structural information by diluting the explicit connectivity, particularly when the graph structure is complex. This thesis makes an attempt to tackle these issues focusing on two major challenges: first, the creation and improvement of neural text generation systems that can better operate when consuming graph-based input data. Second, we examine text-to-text pretrained language models for graph-to-text generation, including multilingual generation, and present possible methods to adapt these models pretrained on natural language to graph-structured data. In the first part of this thesis, we investigate how to directly exploit graph structures for text generation. We develop novel graph-to-text methods with the capability of incorporating the input graph structure into the learned representations, enhancing the quality of the generated text. For AMR-to-text generation, we present a dual encoder, which incorporates different graph neural network methods, to capture complementary perspectives of the AMR graph. Next, we propose a new KG-to-text framework that learns richer contextualized node embeddings, combining global and local node contexts. We thus introduce a parameter-efficient mechanism for inserting the node connections into the Transformer architecture operating with shortest path lengths between nodes, showing strong performance while using considerably fewer parameters. The second part of this thesis focuses on pretrained language models for text generation from graph-based input data. We first examine how encoder-decoder text-to-text pretrained language models perform in various graph-to-text tasks and propose different task-adaptive pretraining strategies for improving their downstream performance. We then propose a novel structure-aware adapter method that allows to directly inject the input graph structure into pretrained models, without updating their parameters and reducing their reliance on specific representations of the graph structure. Finally, we investigate multilingual text generation from AMR structures, developing approaches that can operate in languages beyond English