[Paper Review] Automating Research Synthesis with Domain-Specific Large Language Model Fine-Tuning

Teo Sušnjak, Peter K. Hwang+4 more|arXiv (Cornell University)|Apr 8, 2024

Scientific Computing and Data Management11 citations

TL;DR

This paper demonstrates finetuning open-source LLMs on domain-specific SLR papers to automate the knowledge synthesis phase of systematic literature reviews, addressing hallucination and provenance auditing, and validates the approach by replicating a PRISMA-conforming SLR.

ABSTRACT

This research pioneers the use of fine-tuned Large Language Models (LLMs) to automate Systematic Literature Reviews (SLRs), presenting a significant and novel contribution in integrating AI to enhance academic research methodologies. Our study employed the latest fine-tuning methodologies together with open-sourced LLMs, and demonstrated a practical and efficient approach to automating the final execution stages of an SLR process that involves knowledge synthesis. The results maintained high fidelity in factual accuracy in LLM responses, and were validated through the replication of an existing PRISMA-conforming SLR. Our research proposed solutions for mitigating LLM hallucination and proposed mechanisms for tracking LLM responses to their sources of information, thus demonstrating how this approach can meet the rigorous demands of scholarly research. The findings ultimately confirmed the potential of fine-tuned LLMs in streamlining various labor-intensive processes of conducting literature reviews. Given the potential of this approach and its applicability across all research domains, this foundational study also advocated for updating PRISMA reporting guidelines to incorporate AI-driven processes, ensuring methodological transparency and reliability in future SLRs. This study broadens the appeal of AI-enhanced tools across various academic and research fields, setting a new standard for conducting comprehensive and accurate literature reviews with more efficiency in the face of ever-increasing volumes of academic studies.

Motivation & Objective

• Motivate the use of finetuned LLMs to streamline the knowledge synthesis stage of SLRs.
• Develop methods to convert selected papers into finetuning datasets.
• Mitigate LLM hallucination and enable auditing of model responses to sources.
• Evaluate the framework by replicating a PRISMA-conforming SLR as a gold standard.

Proposed method

• Propose a four-step SLR-automation framework focused on knowledge synthesis.
• Step 2a automates Q&A data extraction and synthesis from selected papers using an LLM.
• Step 2b inserts explicit knowledgeTokens to separate fine-tuned knowledge from pretraining.
• Step 3 applies PEFT-based fine-tuning and optional Retrieval-Augmented Generation (RAG).
• Step 4 verifies factual accuracy with an automated, source-auditable evaluation dataset.

Experimental results

Research questions

• RQ1 RQ1: Can LLMs with finetuning and/or RAG facilitate the synthesis phase of an SLR?
• RQ2 RQ2: How can finetuning datasets be automatically extracted from an SLR corpus?
• RQ3 RQ3: Can finetuning be effectively conducted on relatively small, narrow-domain datasets?
• RQ4 RQ4: Can high fidelity and auditable provenance of LLM outputs be achieved and what metrics are suitable?
• RQ5 RQ5: Are the proposed approaches viable for replicating a published SLR?

Key findings

• The framework aims to automate knowledge synthesis while maintaining factual fidelity and auditable sources.
• PEFT enables efficient fine-tuning by updating a subset of parameters.
• RAG can supplement inference with domain-specific documents to improve factual grounding.
• The approach is demonstrated via a use case that replicates a PRISMA-conforming SLR in learning analytics.
• A Python package is released to facilitate data curation for LLM fine-tuning.
• The study advocates updating PRISMA guidelines to accommodate AI-driven processes.