Developing and Comparing Machine Reasoning Models to Humans in NLP Tasks

Neural Language Models represent a category of computational systems designed to learn task performance directly from raw textual inputs. Their increasing popularity stems from their versatility and remarkable success across diverse domains, such as their transformative impact on machine translation, surpassing traditional machine learning methods. Despite these achievements, a crucial aspect remains unaddressed: the interpretability of the model's decision-making process. Rationale extraction endeavors to furnish explanations that are both faithful (reflective of the model's behavior) and plausible (convincing to humans) by highlighting influential inputs without compromising task model performance. Prior research has primarily focused on optimizing plausibility using human highlights when training rationale extractors, while jointly training the task model to optimize for predictive accuracy and faithfulness. In this thesis, we delve into the significance of explanations, the associated challenges, and the research landscape in this field. We also introduce REFER, a framework that incorporates a differentiable rationale extractor that facilitates back-propagation through the rationale extraction process. Through joint training of the task model and rationale extractor with human highlights, our analysis demonstrates that REFER achieves significantly improved results in terms of faithfulness, plausibility, and downstream task accuracy on both in-distribution and out-of-distribution data compared to previous baselines.