GUPEA

Gothenburg University Publications Electronic Archive

GUPEA is a platform for e-publishing of theses, student essays and other research publications.

Communities in Gupea

Select a community to browse its collections.

Recent Submissions

  • Expert-Owned Validation of Enterprise RAG Assistants: An Exploratory Study of Test-Suite Adequacy and Feedback- Driven Refinement
    (2026-10-12) Adam, Muntasir; Mohammedali, Maab; Göteborgs universitet/Institutionen för data- och informationsteknik; University of Gothenburg/Department of Computer Science and Engineering
    Background: Many organizations are deploying retrieval-augmented generation (RAG) question-answering systems for internal use. However, there is limited empirical guidance on how expert-owned test suites and user feedback can support the structured assessment and improvement of such systems, and help organizations judge when testing is “good enough” in practice. Objective: This thesis investigates how an expert-owned validation setup can be used to evaluate an enterprise RAG-based question-answering assistant in an industrial setting. The study examines how formal test material, expert review, automated harness indicators, and user feedback contribute to judgments of answer acceptability, test-suite adequacy, and trust in practice. An HR knowledge-base chatbot deployment at Getinge serves as an exploratory setting for these broader software-engineering questions. Method: The study followed a Design Science Research approach in an industrial setting. The studied validation artifact consisted of a formal 57-case expert-owned test suite with preferred answers, an expert accept/decline review process, a test harness, user feedback, and a feedback-to-action workflow. The research team designed the study, organized and conducted the empirical evaluation, moderated three focus-group sessions involving six employees and one HR validation expert, analyzed the available expert material and harness export, reviewed 34 logged user-feedback items, and examined project and stakeholder documentation. The employees represented roles or responsibilities related to UX, HR, training, analytics, AI, and product or business development. The concrete setting was a Danish HR knowledge base; the study focused on the validation process rather than on country-specific policy content. Results: The findings indicate that test-suite adequacy depended primarily on the composition of the test material rather than on the number of test cases. Important characteristics included realistic employee phrasing, high-risk and ambiguous scenarios, and cases that exposed weaknesses in the underlying knowledge base. The formal 57-case test suite was executed five times per case, resulting in 285 harness executions. The harness reported an overall Average Quality in Agreement of 72 %, calculated from embedding-based cosine-similarity comparisons between generated answers and expert-defined preferred answers. In parallel, the HR validation expert accepted 55 cases and declined 2 cases, highlighting differences between automated similarity indicators and expert judgments of answer acceptability. Expert review and user feedback also informed candidate test-suite refinements and identified broader improvement needs related to knowledge-base content, interface design, scope communication, and system performance. Conclusion: Based on this exploratory study, the thesis provides practical guidance for the validation and improvement of enterprise RAG assistants. The study suggests that effective validation depends on expert-owned test suites that evolve through expert review and user feedback, with adequacy judged by scenario realism, risk relevance, and coverage of ambiguous cases rather than by test count alone. Automated harness indicators provided descriptive monitoring information, while expert review assessed policy consistency, completeness, grounding, scope, and potential consequences for employees. The thesis also formulates an empirically grounded Expert-Owned RAG Validation Refinement Workflow for reviewing, classifying, and routing observed issues to test-suite refinement, knowledge-base clarification, interface or source-presentation changes, scope communication, or technical-performance action. The findings are exploratory and context-dependent and do not provide a general measure of RAG reliability.
  • Reconstructing the Pre-Industrial Occupational Structure of Swedish Parishes - Data and Methods, 1749–1855
    (2026-10) Chiaiese, Dario; Missiaia, Anna; Sandgren, Fredrik
    This paper reconstructs the occupational structure of 1720 Swedish pseudo-parishes between 1749 and 1855 using the data contained in Tabellverket, a unique collection of population and occupational records covering the country from the mid-eighteenth to the mid-nineteenth century. The original Tabellverket records are characterized by an unsystematic coverage in time and space and across occupational categories, changes in parish and pastorate boundaries, and missing observations that have so far inhibited its use at the parish level. We develop a procedure for transforming these records into a balanced panel data. The resulting dataset provides for the first time parish-level estimates of employment for 11 categories for the entire period 1749-1855 and 43 more detailed categories for the period 1805-1855 for both genders for 27 benchmarks, establishing a new empirical basis for studying long-run structural change in pre-industrial Sweden.
  • Konung Eric och hertig Johan : historiskt skådespel i fem akter /
    (1836) Franul von Weissenthurn, Johanna,
  • Litteraturhistoriens visklek – Författarbilden av Victoria Benedictsson/Ernst Ahlgren
    (2026-10-12) Gunnarsson Rathsman, Matilda; University of Gothenburg/Department of Literature, History of Ideas, and Religion; Göteborgs universitet/Institutionen för litteratur, idéhistoria och religion
    This thesis contains an investigation of how the author portrait of Victoria Benedictsson/Ernst Ahlgren (1850–1888) have been constructed and deconstructed in Swedish literary history. I aim to examine how different variations of the author portrait exists and relates to each other in different times. The historians might create the literary history, but they cannot control the perceptions of their creation. Since the public perception of the authorship consists of more than sources related to literary history, the research material also covers national and local newspaper articles from 1884–2008, a letter correspondence between Axel Lundegård and Albert Bonnier, biographies from 1889–2022, and different perceptions of and opinions on the author’s diary. The analysis is based on Pierre Bourdieus theories of how different agents relates to each other in the literary field depending on their habitus and capital. It also includes a qualitative part with tables of collected data on specific words different agents use to describe Benedictsson/Ahlgren. By combining a quantitative and qualitative research method, the ambition is to expose differences and similarities in the agents’ choices of words. I have identified seven different portraits of Benedictsson/Ahlgren: The Postmaster’s Wife, the Author of Hörby, Lundegård’s Lifeline, the Emotional Being, the Contradictory Human, the Emancipatory Author and the Most Important Female Author of the 1880’s. The result shows that time and context are relevant factors that explains how and why the author portrait changes. Since the time of Benedictsson/Ahlgren’s passing her popularity has continued to increase, especially in Hörby (her place of birth). Scientists, literary critics and biographers all tend to acknowledge the author’s diary as fiction but still read it as a reflection of her reality, in situations when their perceptions correspond with their own agenda. It also shows a difference depending on what name the agent uses; Victoria is associated with (female) weakness while Ernst is associated with (male) strength. Literary scientists have different opinions on whether Axel Lundegård has stewarded or distorted the literary legacy of Benedictsson/Ahlgren. This also explains why different author portraits can be presented by the same kind of agents, at the same time. One aspect the scientists do agree on: Benedictsson/Ahlgren should have a given position in the Swedish literary canon.