Off-campus UMass Amherst users: To download campus access dissertations, please use the following link to log into our proxy server with your UMass Amherst user name and password.

Non-UMass Amherst users: Please talk to your librarian about requesting this dissertation through interlibrary loan.

Dissertations that have an embargo placed on them will not be available to anyone until the embargo expires.

Author ORCID Identifier


Campus-Only Access for One (1) Year

Document Type


Degree Name

Doctor of Philosophy (PhD)

Degree Program

Computer Science

Year Degree Awarded


Month Degree Awarded


First Advisor

Andrew McCallum

Subject Categories

Computer Engineering


Recent years have witnessed impressive progress in question answering thanks in part to advances in neural methods for machine reading comprehension (MRC). Despite their successes, the prediction mechanism of these models are largely opaque to humans, limiting their reliability and maintainability in real-world scenarios. In this thesis, we explore the case-based reasoning approach to improve accuracy, efficiency, interpretability, and controllability of neural machine readers. Case-based reasoning (CBR) works by retrieving past problems similar to the current one and reusing their solutions to address the new problem. For the question-answering setting, we consider a case to be a question-answer pair with its supporting passages. We first approach the task of Machine Reading Comprehension (MRC), in which the answer-supporting passage is already given. We propose a CBR reader with a reuse step that represents solutions of past cases by their contextualized answer embeddings and performs cosine-similarity search over candidate spans of the current context. We show that past cases can be used to infer and interpret new ones, resulting in both fast, accurate inferences and highly interpretable predictions. Second, we extend the proposed CBR reader to the full end-to-end open-domain question answering where the answer-supporting passages must be retrieved. Here we demonstrate significant advantages in the time and space efficiency of our method on open-domain question answering (ODQA), made possible by using pre-computed contextualized token-embeddings in answer passages—all without losing accuracy with respect to traditional parametric machine readers. However, we observe that the precision of the proposed CBR reader deteriorates in the presence of irrelevant passages introduced by the retriever. To address this problem, we will explore representation learning strategies for observed case solutions. We will also study the effects of similar cases on a prediction and describe how a mis-prediction can be fixed by adding cases, without the need to alter parameters. We plan to empirically verify our intuition that by using similar past cases our improved CBR algorithm can achieve high accuracy, fast inference, whilst being more interpretable and controllable.


Creative Commons License

Creative Commons Attribution 4.0 License
This work is licensed under a Creative Commons Attribution 4.0 License.

Available for download on Saturday, February 01, 2025