Integrating knowledge graphs and multilingual scholarly corpora for domain-adaptive LLMs in SSH

2026-07-08 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Research Methodology & Innovation, Social Sciences & Behavioral Studies · Depth: Expert, extended

Summary

The ReSearch_SSH use case, developed within the European LLMs4EU project, aims to adapt Large Language Models (LLMs) for Social Sciences and Humanities (SSH) research. This initiative integrates knowledge graphs and multilingual scholarly corpora to enhance bibliographic discovery and literature synthesis. Utilizing a GraphRAG architecture, the project fine-tunes a multilingual LLM on SSH-specific datasets, including an approximately 3 million-document ISTEX SSH subset (predominantly French, with Italian and English components) and specialized AIUCD/Umanistica Digitale corpora. The system extends the existing ISIDORE platform, supporting tasks like question answering, comparative document analysis, and literature review. Evaluation combines quantitative benchmarking (retrieval, summarization, traceability, hallucination detection) with qualitative assessment by Digital Humanities experts. The project prioritizes compliance with EU regulations such as the AI Act and GDPR, ensuring transparency, traceability, and ethical data handling.

Key takeaway

For research scientists developing domain-adaptive LLMs, prioritize integrating knowledge graphs and multilingual corpora to ensure outputs are traceable and epistemically sound. Your development process should embed legal and ethical compliance from the outset, particularly for sensitive domains like SSH. Engage expert panels for qualitative evaluation to align models with disciplinary practices, moving beyond generic benchmarks. This approach ensures robust, trustworthy, and regulation-compliant generative AI systems.

Key insights

Adapting LLMs for SSH requires domain-specific data, multilingual support, and ethical compliance, integrated into existing research infrastructures.

Principles

• Domain-driven adaptation is crucial for SSH LLMs.
• Hybrid evaluation combines quantitative and expert assessment.
• Compliance must be embedded by design.

Method

The ReSearch_SSH use case employs a GraphRAG architecture, involving domain alignment via continued pre-training on SSH corpora, retrieval optimization using historical ISIDORE queries, and instruction-tuning for generative tasks like literature synthesis.

In practice

• Use GraphRAG for grounded, traceable LLM outputs.
• Incorporate historical user queries for retrieval tuning.
• Use Wikidata for entity disambiguation.

Topics

• Domain-Adaptive LLMs
• Social Sciences & Humanities
• Knowledge Graph RAG
• Multilingual Scholarly Corpora
• Digital Humanities
• AI Act Compliance

Best for: NLP Engineer, AI Scientist, Research Scientist, AI Ethicist

Related on AIssential

• Characterising LLM-Generated Competency Questions: a Cross-Domain Empirical Study using Open and Closed Models — CompCQ framework systematically characterizes LLM-generated Competency Questions across multiple linguistic, structural, and semantic dimensions.
• AI-assisted cultural heritage dissemination: Comparing NMT and glossary-augmented LLM translation in rock art documents — Glossary-augmented LLMs significantly improve terminology accuracy in specialized translation without sacrificing overall quality.
• DySem: Uncovering Dynamic Semantic Components via Multilingual Consensus for Calculating Semantic Textual Similarity — DySem uses multilingual consensus to find dynamic, sample-specific semantic dimensions for improved STS in LLMs.
• Taxonomies, All the Way Down — Taxonomies, a pillar of library science, provide explicit, auditable structure that significantly enhances LLM accuracy and reliability.
• Context Graphs: Free Association, Consistency, Narratives & Truth — LLMs function via free association on narrative graphs, trading consistency for broader, stochastically generated outputs.
• Assessing LLM Reliability on Temporally Recent Open-Domain Questions — LLMs achieve high semantic alignment through paraphrasing, not lexical reproduction, challenging traditional evaluation metrics.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.