Towards Smarter Wills: Information Extraction and Natural Language Inference for Legal Text Understanding in English

Abstract

This dissertation analyzes legal English as a special register of English employed across the legal domain, with specific emphasis on its use in wills, and develops an integrated foundation for transforming natural language wills into executable smart contracts through three interconnected contributions. The first contribution introduces an artificial intelligence-based system that interprets English language wills and computes their expected devolution outcomes. The system architecture and implementation are described in detail, and evaluation on manually annotated test sets shows that it surpasses a baseline model in both accuracy and accountability while maintaining interpretability and traceability. The second contribution extends this framework through a task aware prompt chaining approach for information extraction. The method divides extraction into an initial classification stage and a targeted extraction stage. The classification stage identifies the types of information present in the input and selects the most relevant examples for inclusion in the extraction prompt. Experiments with two different models show improved performance in few-shot settings and reduced token usage. Incorporating this method into the first system enables more detailed information extraction and provides a basis for adding functionality that relies on finer distinctions in the text. The third contribution presents a natural language inference dataset designed to assess the legal validity of will statements. Each validity judgment requires three inputs: a will statement, a law, and the conditions at the testator’s death, resulting in texts that are longer and more complex than those in standard NLI datasets. Eight neural models trained on this dataset achieve over 80 percent macro F1 and accuracy, although group accuracy, a stricter metric evaluating sets of related examples, remains in the mid 80s at best, indicating only superficial task understanding. Ablation and explanation analyses further show that models draw on all three text segments but sometimes rely on semantically irrelevant tokens. While presented as an independent resource, this study can be developed further to extend the first system by supporting automated evaluation of the legal validity of testamentary clauses. Taken together, these studies provide a unified foundation for transforming natural language wills into machine-interpretable smart wills. They also deepen our understanding of the linguistic and structural features of legal documents, including their intricate syntactic patterns, specialized terminology, and reliance on external contextual information such as legal rules and conditions. The findings from this work will support the development of accountable systems for the legal domain and contribute to broader progress toward smarter, more interpretable wills.