Masterarbeit
Knowledge-Graph Based Approaches for Reducing Hallucinations in Large Language Models
Completion
2024/12
Research Area
Intelligent Information Management
Advisers
Description
Transformer-based Large Language Models exhibit impressive capabilities in natural language understanding, processing, and generation. As the model size increases, they demonstrate improved logical reasoning and contextual learning. Leveraging terabytes of data for pre and post-training enhances their comprehension of global events and world knowledge, enabling them to proficiently tackle queries related to worldly matters. However, these models are susceptible to generating inaccuracies, leading to factual errors, or 'hallucinations.' Given the growing influence of AI and language models across industries and their impact on various fields, ensuring their trustworthiness has become crucial.
This thesis aims to investigate different approaches based on Retrieval-Augmented-Generation (RAG) that integrate Knowledge Graphs to mitigate hallucinations in Large Language Models. The research involves analyzing various existing methodologies and developing a demonstrator based on RAG and Knowledge Graphs. Evaluation will be conducted using benchmarks such as TruthfulQA to assess the effectiveness of the proposed method.
