From Unstructured Text to Executable Knowledge: Automated Transformation of High-Stakes Domain Documents
TL;DR
This thesis investigates how unstructured documents can be reliably transformed into structured, machine-actionable, and where appropriate, executable representations using small, open-source language models deployable in-house within an organization's own private infrastructure, avoiding external proprietary cloud services.
Abstract
High-stakes domains such as finance, insurance, and law rely on long, specialized, and often confidential documents that are expensive to analyze manually and risky to send to proprietary cloud Large Language Models (LLMs). This thesis investigates how such unstructured documents can be reliably transformed into structured, machine-actionable, and where appropriate, executable representations using small, open-source language models deployable in-house within an organization's own private infrastructure, avoiding external proprietary cloud services. The work develops a progression from schema-guided extraction, through downstream execution of the resulting structured artifacts, to schema-free settings where the structure itself must be learned from the data. Financial contracts serve as the primary experimental testbed given their richness and complexity, validating a broader hypothesis: that carefully structured pipelines with small open-source LLMs can deliver reliable and cost-effective knowledge extraction from complex documents, while providing provenance and auditability. This dissertation advances the thesis that small open-source LLMs can support accurate, auditable, and executable knowledge transformation when their role is restricted to bounded, evidence-grounded decisions within a deterministic structural pipeline.
