Can Structured Data Reduce Epistemic Uncertainty?
Shriram M S, Sushmitha S, Gayathri K S, Shahina A
TL;DR
The paper tackles reducing epistemic uncertainty ($EU$) in large language models by leveraging structured data through ontology alignment, formalizing the data as $ \{(x_i,y_i)\}_{i=1}^n \in (\mathcal{X}\times\mathcal{Y})^n$ and a predictor $F:\mathcal{X}\to\mathcal{Y}$. It introduces a framework that aligns two ontologies using a transformer to obtain equivalence ($c\equiv c'$) and subsumption ($c_{s}\sqsubseteq c$) mappings, and then integrates these subsumptions into Retrieval-Augmented Generation prompts to improve context and factuality. In medical-domain experiments, the approach yields an 8.97% increase in contextual similarity, a 1% rise in factual accuracy, and a 4.847% reduction in the hallucination index. This work provides a concrete path to reduce epistemic uncertainty in LLMs via structured data and targeted prompt design, with potential extensions to multi-modal settings and other domains.
Abstract
In this work, we present a framework that utilizes ontology alignment to improve the learning process of deep learning models. With this approach we show that models fine-tuned using ontologies learn a downstream task at a higher rate with better performance on a sequential classification task compared to the native version of the model. Additionally, we extend our work to showcase how subsumption mappings retrieved during the process of ontology alignment can help enhance Retrieval-Augmented Generation in Large Language Models. The results show that the responses obtained by using subsumption mappings show an increase of 8.97% in contextual similarity and a 1% increase in factual accuracy. We also use these scores to define our Hallucination Index and show that this approach reduces hallucination in LLMs by 4.847%.