Domain Adaptation of Foundation Language Models for Legal Document Generation: A Systematic Review and Implications for Analogous Regulated Institutional Contexts

Authors:
Open Access
Publisher website
Google Scholar
Abstract
The adaptation of foundation language models for legal document generation has become a relevant problem in Legal Natural Language Processing, particularly because generated texts must preserve factuality, normative coherence, traceability and control over sensitive information. This systematic literature review maps, classifies and critically analyzes primary studies on domain adaptation techniques for foundation language models applied to legal and normative-argumentative document generation. It also examines the implications of the resulting evidence for analogous regulated institutional contexts that share requirements related to reliability, privacy, auditability and institutional control. The review followed guidelines for systematic reviews in Software Engineering and the applicable items of PRISMA 2020. Searches were conducted in Scopus, Web of Science, IEEE Xplore and ACM Digital Library, covering studies published from January 2021 to December 2025 and complemented by backward snowballing. From 200 initial records, 32 primary studies were included. The corpus is predominantly concentrated in the legal domain, with only one study directly addressing document production in Public Security. The results show a recurrent use of hybrid adaptation strategies involving parameter-efficient fine-tuning, retrieval-augmented generation and structured inference, although the literature does not establish a universally superior combination. Data engineering is marked by a trade-off between scale and control, with frequent use of semi-synthetic datasets. Traditional Natural Language Processing metrics remain useful as initial indicators, but are insufficient to assess factuality, normative validity and argumentative consistency. Ethical safeguards, privacy controls and bias mitigation metrics remain underdeveloped. The findings indicate that the institutional adoption of foundation language models for legal document generation should be treated as a problem of architecture, evaluation and governance, rather than only textual performance. These conclusions are directly grounded in the Legal NLP literature, while their transfer to analogous regulated institutional contexts should be understood as a functionally motivated implication requiring domain-specific empirical validation.

This publication has 34 references indexed in Scilit: