On Knowledge in AI: Epistemic and Ethical Limitations of Language Models and Knowledge Graphs

Abstract

This thesis critically investigates whether or not AI-based knowledge technology built on language models, knowledge graphs, and/or knowledge-enhanced language models deserves the epistemic authority it happens to receive and analyzes its epistemic and ethical "goodness". To this end, this thesis utilizes computer science approaches and philosophical analysis and discusses technical features, as well as engineering and research practices in the field of AI by drawing from feminist epistemological accounts, in particular. The core of this cumulative dissertation comprises three articles that address the following sets of questions: (RQ1) What types of social bias are embedded in knowledge graphs? How are they measured? And what do we know about their causes? (RQ2) Can knowledge enhancement make language models less biased with regards to their knowledge content? Can it help to make language models more objective? (RQ3) How are the measures created that are used to determine a language model's accuracy in reproducing knowledge? How is the quality and representativeness of these measures? This thesis finds that AI-based knowledge technology has several epistemically and ethically problematic characteristics, which cannot be solved solely through technological means. AI development and evaluation must be conducted in a contextualized manner and account for the situatedness of knowledge processes. In drawing from feminist epistemologies, this thesis argues that the AI community must promote emancipatory values and foreground marginalized standpoints to facilitate epistemically and ethically better systems.