ABSTRACT 

Constructing a persuasive legal argument requires identifying precedents that are both factually relevant and legally authoritative, a task made difficult by the scale, length, and technicality of judicial texts. Modern information retrieval systems, including dense retrievers, are effective at modeling semantic similarity but often fail to capture the hierarchical constraints that determine whether a case is actually binding. This dissertation characterizes that failure mode as the Textual Similarity Trap: retrieval models over-prioritize lexical or semantic overlap, causing non-binding but factually similar cases to outrank binding precedents that are more legally relevant yet share less case-specific vocabulary with the query. We argue that effective legal retrieval must model two dimensions of relevance jointly: Material Relevance and Legal Authority. To address this problem, we introduce AuthorityRank, a dual-component neural ranking framework that separates the modeling of textual content from the modeling of institutional authority. The dissertation makes three contributions. First, it isolates Material Relevance to examine the effects of formulaic legal language. Experiments on case law from the Court of Justice of the European Union show that BM25 often outperforms off-the-shelf dense retrievers, although domain-specific fine-tuning improves dense retrieval. Both approaches benefit from repetitive language, but lexical methods remain stronger on longer, more nuanced queries with less verbatim overlap. Second, the dissertation addresses Legal Authority by formulating judicial influence as a link-weight prediction task. Experiments on an inter-court citation network show that adding court metadata, including jurisdiction and court level, to topological features improves the prediction of authority relations and yields better court representations. Third, we show that the AuthorityRank framework, which combines these components, improves retrieval over strong lexical and semantic baselines. Its context-aware authority component models authority across jurisdictions, while its citation-enhanced textual similarity component reduces noise from formulaic legal language by leveraging citation context. Experiments on a novel dataset of U.S. federal case law show the benefit of jointly modeling legal authority and material relevance. The dissertation provides a framework for legal retrieval systems grounded in binding precedent rather than mere topical similarity.

BIOGRAPHY

Larissa Mori is a Ph.D. Candidate in the Edwardson School of Industrial Engineering at Purdue University, with a concentration in Computational Science and Engineering. Originally from Brazil, she holds a Bachelor of Laws (L.L.B.) and a B.A. in Economics from the University of Braslia. Her research focuses on machine learning, with applications in recommendation systems, information retrieval, and personalization in the legal domain. Her work has been published in peer-reviewed journals and conference proceedings, including Entropy, the Natural Legal Language Processing (NLLP) Workshop at EMNLP, and the International Conference on AI and Law (ICAIL), and featured at the Bloomberg Law Symposium on Law, Language, and AI.