Facilitating assessment of topic modelling credibility

Abstract

This thesis investigates the factors influencing the perceived credibility of the Latent Dirichlet Allocation (LDA) algorithm, particularly focusing on the roles of transparency and interface design. Despite LDA’s widespread use in topic modeling within natural language processing, its opaque nature often leads to user distrust, impacting its application in areas requiring reliable decision-making support. The study employs a qualitative research approach, integrating user-based experiments and semi-structured interviews to explore how different levels of transparency and interface designs affect user perceptions of credibility. Experiments were designed to vary in the amount of algorithmic workings exposed to participants, alongside interface variations to assess their impact on user trust and understanding. Results indicate that while increased transparency generally enhances trust, overly complex information presentation can diminish the perceived credibility of the LDA algorithm. Moreover, intuitive and well-designed interfaces were found to significantly bolster user trust and comprehension, suggesting a delicate balance between transparency and usability is crucial for enhancing the credibility of algorithmic systems. This research contributes to the fields of human-computer interaction and machine learning by providing empirical insights into designing more effective user interfaces for complex algorithms. It highlights the importance of user-centered design principles in improving the accessibility and trustworthiness of machine learning models, with implications for developers and researchers seeking to foster greater acceptance and use of AI technologies.