Explain and Improve Natural Language Processing Systems with Human Insights
Zhao, Xinyan
2022
Abstract
Human insights play an essential role in artificial intelligence (AI) systems as it increases the coherence between the human mind and system decisions by allowing the systems to be interpretable or aided by human input. However, the use of human insights in AI systems remains mysterious, despite the fact that their predictive power has demonstrated high correspondence regarding system output. This missing bridge between human and AI is caused mainly by the black-box nature of deep learning-based models. It often presents two challenges in applying these systems to critical domains such as healthcare, finance, or law practice. First, the low coherence between humans and AI often leads to systems with high performance yet lack human trust. Interpreting a system decision demands simple, faithful, and consistent explanations for humans to understand. Failing to provide such explanations decreases human trust in the system and results in serious outcomes. The second challenge concerns the large-scale data annotation that is usually expensive and impractical in the above critical domains as it requires specially trained domain expertise. To advance the AI applications in critical domains, in this dissertation, I explore methods that increase the coherence between humans and AI while maintaining high system correspondence. I look at AI systems from the ``pipeline'' (instead of ``model-only'') perspective and focus on designing ``grey-box'' pipelines that have not only high system correspondence but also high coherence between humans and the pipeline. I investigate two directions. The first direction investigates if we could improve the human-AI coherence by mimicking human behavior with high-correspondence black-box models. The second direction investigates if mimicking human insights could further improve system correspondence. To investigate the first direction, I evaluate how black-box models could mimic human insights as explanations in various formats, including natural language, extractive data snippets, and human heuristics. In Chapter 3, I introduce an approach that generates natural language explanations based on human rationales that improve the explanation quality. Furthermore, in Chapter 4, I conduct a study in a clinical scenario where the system predicts the medication use of patients with inflammatory bowel disease (IBD) based on patients' medical records. I show that sentence-based explanations could be well predicted by using heuristics-based weak supervision with minimal annotation or annotation-based supervised learning with large annotation budgets. Lastly, in Chapter 5, I examine if black-box models could well mimic human heuristics. Using adverse drug event identification as an example, I show that while human heuristics could be easily implemented into interpretable rule-based or white-box systems, they are often over-simplified. To overcome this limitation, I show that the semantics of heuristics could be well augmented by the black-box model with a light-loaded annotation effort from human-in-the-loop. To investigate the second direction, I further evaluate if human insights could improve system correspondence. In Chapter 3, I show that high-quality language explanations bring system correspondence gain. In Chapter 4, I show that identifying evidence support from patient notes brings substantial performance gain for the inference of IBD medication history. In Chapter 5, I show that the semantically augmented human heuristics could be easily applied to weak supervision and generate strong correspondence. This dissertation contributes to the field of explainable AI. More specifically, this dissertation provides opportunities for designing AI systems that desire explanations to increase human trust in critical domains by collaborating with human insights.Deep Blue DOI
Subjects
natural language processing clinical natural language processing explainable AI human-centered AI
Types
Thesis
Metadata
Show full item recordCollections
Remediation of Harmful Language
The University of Michigan Library aims to describe library materials in a way that respects the people and communities who create, use, and are represented in our collections. Report harmful or offensive language in catalog records, finding aids, or elsewhere in our collections anonymously through our metadata feedback form. More information at Remediation of Harmful Language.
Accessibility
If you are unable to use this file in its current format, please select the Contact Us link and we can modify it to make it more accessible to you.