Masterarbeit
Content Analysis using LLMs for Trust Evaluations by ConTED
Completion
2024/10
Research Area
Students
Borneel Bikash Phukan
Advisers
Description
The web as we know today is dominated by web applications connected with databases and services, most of which are owned by large technology corporations. These corporations mostly use centralized web infrastructures and ecosystems, which raises data and security concerns. Users have no or limited rights to manage and control their own data, which increases the need for a decentralized web where data and applications are distributed across a wide network of independent host systems. However, trust may not be given for external stored data, which can also be misleading, harmful, or malicious. The trust framework ConTED for decentralized web applications is a candidate to enable trust awareness with respect to external data stores in the decentralized web. ConTED has three main components for trust evaluation: Content Analysis, Context Analysis, and Behavioral Analysis. The component of Content Analysis focuses on calculating Specificity and Likelihood, as well as Topic Classification. Specificity measures how relevant and focused a piece of information is to the task at hand, while Likelihood assesses the probability of the information being accurate.
This thesis addresses the issue of trust evaluation by proposing a method for assessing trustworthiness within the ConTED framework. It focuses thereby on the analyzation how to use Large-Language-Models (LLMs) to analyse the content of incoming data from third-party data stores. This includes the preprocessing of RDF data as well as the inclusion of LLMs to solve the three tasks required by ConTED's content analysis.
The objective of this master's thesis is to find an approach or combination of approaches to solve the previously mentioned problem in the context of content analysis of linked data for trust awareness via ConTED making use of LLMs. This particularly includes the state of the art regarding Trust Awareness and content-based analysis of linked data using LLMs. The demonstration of feasibility with an implementation demonstrator of the concept is part of this thesis as well as a suitable evaluation with exemplary use cases.
