A Survey on Large Language Models for Critical Societal Domains: Finance, Healthcare, and Law

In one sentence

Summary

The paper addresses the critical need to understand how large language models are transforming high-stakes societal domains: finance, healthcare, and law. It positions LLMs as transformative tools that require specialized handling due to the domains' reliance on professional expertise, confidential data, multimodal documents, strict regulatory compliance, and the necessity for explainability and fairness. The survey aims to unify the fragmented literature by offering a structured overview of methodologies, applications, and ethical considerations across these three sectors, highlighting their shared challenges and unique requirements. The methodology involves a systematic categorization of existing literature into tasks, datasets, model architectures, and evaluation metrics. In finance, it details tasks such as sentiment analysis, information extraction, question answering, and stock movement prediction, reviewing models like FinBERT, BloombergGPT, and FinMA. The paper analyzes training paradigms, including pre-training on financial corpora and instruction fine-tuning. It also covers healthcare tasks like diagnostic support and legal tasks like contract review, emphasizing the shift from general NLP to domain-specific LLMs. Experiments cited include performance comparisons on benchmarks like FinQA and TAT-QA, and evaluations of proprietary models like BloombergGPT against open-source baselines. Key findings indicate that while LLMs show promise in enhancing financial analytics and legal interpretation, they face significant hurdles in numerical reasoning, hallucination, and regulatory compliance. The survey identifies under-explored areas such as financial fraud detection and risk assessment. It concludes that future research must prioritize interdisciplinary cooperation, robust ethical frameworks, and methodological advancements to ensure transparent, fair, and robust AI systems. The paper provides a curated reading list and emphasizes the importance of addressing data privacy and bias to mitigate risks in precision-dependent sectors.

Research problem

The paper investigates the integration of large language models into critical societal domains: finance, healthcare, and law. It addresses the challenge of adapting general LLMs to these high-stakes fields, which demand professional expertise, handle confidential data, process multimodal documents, and adhere to strict regulations. The core problem is to synthesize existing research to identify effective methodologies, evaluate performance, and highlight ethical and practical barriers to deployment.

Core contributions

• Provides the first comprehensive survey of LLMs across finance, healthcare, and law, highlighting shared and unique challenges.
• Categorizes and reviews financial NLP tasks, datasets, and specialized models like FinBERT and BloombergGPT.
• Analyzes training paradigms, including pre-training, instruction fine-tuning, and their impact on domain-specific performance.
• Identifies under-explored financial tasks such as fraud detection and risk management, suggesting future research directions.
• Discusses ethical concerns and regulatory compliance requirements specific to high-stakes domains.
• Offers a structured taxonomy and a curated reading list to facilitate future interdisciplinary research.

Method

• Systematic literature review of LLM applications in finance, healthcare, and law.
• Categorization of tasks into sentiment analysis, information extraction, question answering, and stock prediction.
• Analysis of model training paradigms: pre-training, instruction fine-tuning, and domain-specific adaptation.
• Evaluation of model performance on standard benchmarks and proprietary datasets.
• Discussion of ethical frameworks and regulatory compliance requirements.
• Identification of gaps in current research, such as fraud detection and risk management.

Results & findings

• BloombergGPT outperforms open-domain LLMs on financial QA and sentiment tasks but faces NER limitations.
• FinMA-30B surpasses BloombergGPT in classification tasks but lags in numerical reasoning compared to GPT-4.
• Instruction fine-tuning improves performance on diverse financial tasks but requires high-quality reasoning data.
• Financial sentiment analysis and information extraction are well-established, while fraud detection remains under-explored.
• Multimodal integration is crucial for healthcare and legal applications but presents significant technical challenges.
• Ethical concerns like bias, hallucination, and data privacy are critical barriers to deployment in regulated industries.

Data & experiments

• Reviews datasets for financial sentiment analysis, information extraction, question answering, and stock movement prediction.
• Analyzes performance of models like FinBERT, BloombergGPT, and FinMA on benchmarks such as FinQA, TAT-QA, and FiQA.
• Evaluates instruction-tuned models against proprietary baselines in classification and reasoning tasks.
• Covers healthcare and legal datasets and tasks, though with less depth than the finance section.
• Discusses multimodal data challenges in medical imaging and legal evidence processing.

Limitations & caveats

• The survey is broad and may not provide deep technical details for specific sub-tasks.
• Focuses on recent developments, potentially missing earlier foundational work in financial NLP.
• Ethical discussions are general and may not cover all jurisdiction-specific regulations.
• Performance comparisons are based on reported results, which may vary due to different evaluation settings.
• The scope is limited to finance, healthcare, and law, excluding other critical domains.

Curator notes

Discovery status: round2_promoted. Source links: Advancing Financial Engineering with Foundation Models: Progress, Applications, and Challenges; FinChain: A Symbolic Benchmark for Verifiable Chain-of-Thought Financial Reasoning; Standard Benchmarks Fail -- Auditing LLM Agents in Finance Must Prioritize Risk