The InterDev Framework: Harnessing Semantic Web Technologies to Enable FAIR Evidence in International Development

Abstract

International development research depends on the ability to discover, compare, and synthesise evidence from heterogeneous impact evaluations, systematic reviews, registries, and contextual datasets. In practice, however, this evidence remains fragmented across repositories, reporting formats, and levels of abstraction, making it difficult for researchers and practitioners to reuse findings efficiently or transparently. This thesis addresses that problem through the design, implementation, and evaluation of InterDev, an ontology-driven linked-data platform for managing and exploring international development evaluation evidence. InterDev is built around an ERCT ontology-aligned knowledge graph that supports evidence browsing, faceted filtering, collection-based curation, structured submission, and export, without requiring users to possess semantic web expertise. The research follows a six-phase User-Centred Design process. Initial requirements were derived through a domain-grounding study based on a review of systematic reviews in international development, which identified recurring needs for advanced filtering and contextualisation, access to both granular and aggregated evidence representations, and support for cross-study comparison and harmonisation. The thesis then reports an iterative evaluation programme across three design iterations. The first iteration established baseline usability and identified major interaction breakdowns in filtering, collection management, export, and manual submission. The second iteration introduced multi-criteria filtering, multi-collection workflows, improved export, and LLM-assisted submission. This substantially reduced manual contribution effort and improved completion rates relative to manual entry, while remaining compatible with human review and validation. The third iteration extended evaluation beyond usability to examine the epistemic effects of LLM-mediated exploration under different provenance conditions. Results show that viii conversational mediation reduced perceived workload, but did not reliably improve task performance or efficiency once task difficulty was controlled. Surfaced provenance changed how users judged outputs and influenced confidence and verification behaviour, but did not by itself improve trust or confidence calibration. To probe whether these interaction-level trade-offs were specific to international development or reflected broader properties of exploratory AI systems, the thesis also reports a complementary cross-domain evaluation in a digital-library setting. That study found comparable patterns: conversational interfaces reduced interactional effort and often increased user confidence, but did not eliminate the need for verification and careful synthesis. This cross-domain result is used as bounded supporting evidence for the broader relevance of the interaction design implications, without claiming generalisation of InterDev itself. Together, the findings show that ontology-driven linked-data platforms are a viable basis for reusable evidence infrastructures in international development, and that LLM mediation can reduce interactional effort, particularly for contribution and exploratory access. However, the thesis also demonstrates that usability gains do not automatically translate into improved epistemic outcomes. The central design challenge is therefore not simply adding AI assistance, but integrating it in ways that preserve provenance, support verification, and enable calibrated, reliable evidence use.