AI-driven Discovery Systems for Multilingual Academic Libraries

Thesis Abstract

The present study investigates how AI-driven discovery systems can enhance access to multilingual resources in academic libraries, addressing persistent barriers related to language diversity, metadata fragmentation, and semantic gaps that hinder discoverability for non-English users. Despite rapid advances in natural language processing and machine learning, a gap remains in empirically assessing how integrated AI-based discovery interfaces perform across languages, disciplines, and user groups, and how such systems influence research workflows and information equity. The aim is to design, implement, and evaluate an AI-enhanced discovery platform that supports multilingual querying, cross-lingual relevance ranking, and culturally aware faceting, thereby improving retrieval effectiveness and user satisfaction for diverse scholarly communities. The study sets four specific objectives (1) to develop an AI-assisted discovery prototype that supports multilingual indexing, cross-lingual search, and adaptive relevance ranking using transformer-based embeddings; (2) to evaluate the prototype’s information retrieval performance against a baseline federated search across English, Spanish, Mandarin, Arabic, and French corpora totaling approximately 4 million metadata records and 1.2 million full-text documents; (3) to examine user experiences, cognitive load, and satisfaction across multilingual users using mixed methods and validated instruments; and (4) to formulate guidelines for implementing equitable discovery systems in multilingual academic libraries. The research employs a mixed-methods design grounded in the relevance theory and socio-technical systems theory to capture both technical performance and user-centered outcomes. The primary population includes library users and information professionals in three large research libraries within a metropolitan university consortium, with a purposive sample of 360 undergraduate and postgraduate researchers, 40 subject librarians, and 15 system administrators. Instruments comprise an AI-driven discovery prototype interface, a standard search benchmarking corpus, system usability scale (SUS) questionnaires, a multilingual information retrieval acceptance model (MIAM) survey, semi-structured interviews, and log data captured over a 12-week evaluation period. Validity and reliability are established through triangulation of quantitative metrics (precision, recall, F1, and mean reciprocal rank), A/B testing of search variants, and qualitative coding using thematic analysis. Data analysis employs a combination of regression analysis to identify predictors of search satisfaction, ANOVA to compare performance across languages and disciplines, and hierarchical linear modeling to account for user-type effects. Topic modeling via Latent Dirichlet Allocation (LDA) and sentence-transformer embeddings are used to analyze query expansion effectiveness and cross-lingual retrieval quality. A conceptual model integrating multilingual ontology alignment and user-contextualization feedback is iteratively refined through empirical results. Expected findings include (a) improved precision and recall in multilingual retrieval due to cross-lingual embeddings and ontology-aligned facets; (b) higher task success rates and reduced search effort for non-English-speaking users, evidenced by SUS scores above 75 and reduction in search time by 20–30% in comparison with baseline; (c) enhanced perceived relevance and satisfaction through adaptive ranking that considers language proficiency, topic familiarity, and user intent; (d) identification of bottlenecks in metadata quality and supply chain delays for multilingual records, and (e) evidence that explainable AI features (natural language explanations for re-ranked results) increase trust and adoption among users. The study contributes to knowledge by providing empirical evidence on the design and evaluation of AI-driven, multilingual discovery systems in higher education libraries, detailing the interplay between technical components (cross-lingual embeddings, multilingual metadata normalization, and explainability) and user-centered outcomes (usability, satisfaction, and equitable access). Theoretical contributions include operationalization of relevance and language-proximal search satisfaction within a socio-technical lens, and validation of a deployment-ready framework for multilingual discovery that can be generalized to diverse library ecosystems. Based on the findings, practical recommendations will address metadata standardization, cultural and linguistic relevancy tuning, human-in-the-loop governance for AI components, and scalable deployment strategies. The study concludes that AI-driven discovery systems can meaningfully reduce language barriers and enhance research equity in academic libraries when designed with robust multilingual NLP, transparent ranking, and continuous user-centered evaluation.