Abstract
This thesis addresses the challenge of enhancing the reasoning capabilities of natural language processing (NLP) models. Recently, transformer-based pre-trained language models have shown a lot of progress in various NLP tasks. However, the reasoning abilities of these models remain limited, often relying on spurious pattern recognition rather than deep understanding. This research aims to strengthen these capabilities across various NLP tasks, leveraging novel techniques and methodologies. We begin by exploring the task of domain adaptation using commonsense knowledge graphs, demonstrating that these graphs can facilitate significant performance improvements in tasks across unseen domains. Our subsequent chapters are focused on improving and formulating various reasoning tasks over dialogues. We showcase enhanced performance in emotion recognition in conversation tasks through reasoning-augmented language models. We also construct a new corpus for dialogue explanation and reasoning, revealing the constrained performance of transformer-based models on dialogue-level tasks. Further, we introduce another dataset focused on utterance-level contextualized commonsense inference, emphasizing the importance of dialogue-driven commonsense knowledge. A novel framework employing graph neural networks and topological sorting is also proposed to enhance document-level coherence reasoning, used for the task of sentence ordering. We also investigate reasoning-based multi-choice question-answering tasks, for which we present a new objective function that significantly improves language model performance across various reasoning tasks. Lastly, our work on multimodal reasoning, particularly visual question answering (VQA), introduces language-based guidance to enhance vision-language models, illustrating notable advancements in knowledge-augmented and scientific VQA tasks. This study advances the understanding of natural language reasoning research topics and also sets new state-of-the-art results across a broad range of natural language processing tasks.