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

Technical Summary: Integrating Knowledge Graphs and Multilingual Scholarly Corpora for Domain-Adaptive LLMs in SSH

Problem Statement
The integration of Large Language Models (LLMs) into scientific research workflows presents significant methodological and epistemic challenges for the Social Sciences and Humanities (SSH). Current AI-enhanced discovery platforms (e.g., Semantic Scholar, Elicit) are predominantly trained on English-language corpora and prioritize citation-based metrics typical of "hard" sciences. This design reinforces specific models of scientific production that marginalize SSH practices, which rely heavily on multilingualism, book-based scholarship, local research traditions, and qualitative evaluation. Furthermore, SSH research often requires the interpretation of texts within specific historical and theoretical contexts, a nuance that simple topic-matching algorithms fail to capture. Existing infrastructures, while rich in resources, often lack the generative capabilities to synthesize literature or support complex, cross-lingual queries while maintaining the traceability and epistemic responsibility required by humanities scholars.

Methodology
The paper presents ReSearch_SSH, a use case within the European LLMs4EU project and the ALT-EDIC infrastructure, designed to adapt foundation models to SSH research practices. The system employs a Knowledge-Graph-based Retrieval-Augmented Generation (GraphRAG) architecture, structured around four main stages:

1. Data Curation and Domain Alignment:

   • Corpus: The primary alignment corpus is the SSH subset of the ISTEX database (~3 million documents, 60+ languages), supplemented by targeted Italian and bilingual Digital Humanities (DH) datasets (AIUCD conference proceedings and the Umanistica Digitale journal). Additional data includes academic blog posts from Hypotheses and research data from Nakala.
   • Languages: The corpus is multilingual, with a significant focus on French (the dominant non-English component for the ISIDORE ecosystem), Italian, and English.
   • Pre-training: The base model undergoes continued pre-training on these SSH corpora to strengthen linguistic and terminological competence while preserving general reasoning capabilities.

2. Retrieval Optimization (RAG):

   • Retrieval Layer: A domain-adapted retrieval model is fine-tuned on historical query–interaction data from the ISIDORE platform. This interaction-aware strategy promotes documents actually clicked and consulted by scholars, aligning ranking with observed scholarly practices rather than relying solely on textual similarity.
   • Graph Enrichment: Contextualization relies on Wikidata for entity disambiguation and semantic expansion. Relational structures derived from metadata (authors, themes, institutions, citations) and document-level semantic graphs support structured navigation and fine-grained comparison.

3. Generative Task Adaptation:

   • Instruction Tuning: The generative component is instruction-tuned for research-oriented tasks, specifically the construction of state-of-the-art overviews and literature syntheses.
   • Synthetic Data: Fine-tuning data for query understanding is generated by comparing commercial models (e.g., GPT-5) with open European models (e.g., EuroLLM). The project prioritizes open European models to ensure sovereign and transparent AI infrastructures.
   • Output: The system generates structured overviews grounded in explicitly retrieved sources, ensuring transparency and traceability.

4. Evaluation and Compliance:

   • Framework: Evaluation follows the LLMs4EU protocol, involving an independent external team. It includes quantitative benchmarking (retrieval, summarization, traceability, hallucination detection) and qualitative assessment by a panel of Digital Humanities experts from France and Italy.
   • Legal Compliance: The system is designed with "compliance-by-design," adhering to the EU AI Act, GDPR, and copyright regulations. It operates as a downstream modification of a General-Purpose AI (GPAI) model within a controlled research sandbox, avoiding high-risk classification by excluding automated decision-making. Personal data is minimized, and licensed full texts are processed under strict contractual conditions.

Key Contributions

• Domain-Specific Adaptation: The paper proposes a workflow for adapting LLMs specifically to the epistemic and methodological requirements of SSH, moving beyond generic conversational interfaces.
• GraphRAG Architecture: It demonstrates the integration of Knowledge Graphs (Wikidata, OpenAIRE, ISIDORE metadata) with RAG to enhance source grounding, reduce hallucinations, and enable complex relational reasoning across authors and themes.
• Multilingual and Interdisciplinary Focus: The approach explicitly addresses the plurality of scholarly outputs (monographs, datasets, blogs) and the necessity of multilingual support (French, Italian, English) within European research infrastructures.
• Participatory Evaluation: The establishment of an expert panel of SSH scholars ensures that the system's relevance, factual grounding, and usability are assessed against disciplinary standards rather than generic benchmarks.
• Regulatory Alignment: The project integrates legal and ethical constraints (transparency, traceability, data protection) as structural components of the system architecture rather than post-hoc constraints.

Results
The paper reports on an ongoing use case currently in an advanced preparatory phase. Consequently, it does not present final quantitative results or experimental outcomes.

• Current Status: The multilingual corpus is being consolidated, metadata is being normalized, and the fine-tuning strategy is being finalized.
• Immediate Next Steps: The project is scheduled to begin initial experiments focusing on the controlled adaptation of an open foundation model to SSH discourse and the optimization of the KG-RAG pipeline.
• Validation: The evaluation framework is under development, with the construction of ad hoc benchmark datasets tailored to SSH research scenarios.

Significance
The paper positions ReSearch_SSH as a critical step toward developing multilingual, domain-sensitive, and regulation-compliant generative systems that are both technically feasible and methodologically desirable for the SSH community.

• Epistemic Responsibility: By embedding model adaptation within research infrastructures (ISIDORE) and a structured legal framework, the project explores how generative AI can support scholarship while preserving reliability and epistemic responsibility.
• Infrastructure Transferability: While currently anchored to ISIDORE, the modular design of the architecture (separating retrieval, graph enrichment, and generation) is intended to facilitate transferability to other institutional settings with varying metadata quality.
• Community Engagement: The project aims to demonstrate that involving Digital Humanities communities in the development and evaluation of LLMs is essential for creating tools that reflect disciplinary practices and avoid imposing external models of relevance.
• European Sovereignty: By prioritizing open European models and open data infrastructures, the work contributes to the broader goal of developing sovereign, transparent, and value-aligned AI technologies for the European Union.