Enhancing Explainable RAG through Neural-Hungarian Optimization : Hybrid Framework for Transparent Question Answering
Ramanayake, Imanda (2025)
2025
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2025-10-24
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025102410091
https://urn.fi/URN:NBN:fi:tuni-2025102410091
Tiivistelmä
Large language models (LLMs) have profoundly transformed the relationship between artifi-cial intelligence and human - computer interaction. LLMs have ability to produce logical, contex-tually aware, and creative responses and this has made revolutionary advancements possible in the fields of dialogue systems, creating content, and natural language understanding. One sig-nificant advancement among them is Retrieval-Augmented Generation (RAG), which enables models to dynamically access and make decisions from external information sources. RAG combines retrieval-based grounding and generative reasoning to reduce hallucinations and im-prove factual accuracy.
RAG still faces significant obstacles regardless of its achievements. It becomes challenging to fully understand model reasoning due to the complexity created by the integration of retrieval and generation. This makes it difficult for users to understand how retrieved passages impact the model's output. This lack of transparency has an impact on interpretability, accountability, and user trust specially in the areas where factual accuracy is important.
With the objective of improving the performance and interpretability of RAG systems, this thesis presents a novel hybrid explainability framework. The suggested approach uses a novel mechanism known as Neural-Hungarian Optimization to combine combinatorial optimization and neural attention modelling. The approach creates an interpretable ranking of retrieved passages by combining semantic relevance and inter-passage attention coherence into a combined score. After that, these sections are properly arranged using the Hungarian method, which also reduces positional bias (the "lost in the middle" effect) and provides a structured explanation for each generation step.
Experimental results on the WikiQA dataset show that the proposed approach enhances in-terpretability and overall performance. This hybrid approach improves the performance of the test queries by 90 percent and generates clear, quantitative data about how each passage in-fluences the model's selection. In the process of developing high-performing and easily com-prehensible Retrieval-Augmented Generation systems, this work is an important milestone.
RAG still faces significant obstacles regardless of its achievements. It becomes challenging to fully understand model reasoning due to the complexity created by the integration of retrieval and generation. This makes it difficult for users to understand how retrieved passages impact the model's output. This lack of transparency has an impact on interpretability, accountability, and user trust specially in the areas where factual accuracy is important.
With the objective of improving the performance and interpretability of RAG systems, this thesis presents a novel hybrid explainability framework. The suggested approach uses a novel mechanism known as Neural-Hungarian Optimization to combine combinatorial optimization and neural attention modelling. The approach creates an interpretable ranking of retrieved passages by combining semantic relevance and inter-passage attention coherence into a combined score. After that, these sections are properly arranged using the Hungarian method, which also reduces positional bias (the "lost in the middle" effect) and provides a structured explanation for each generation step.
Experimental results on the WikiQA dataset show that the proposed approach enhances in-terpretability and overall performance. This hybrid approach improves the performance of the test queries by 90 percent and generates clear, quantitative data about how each passage in-fluences the model's selection. In the process of developing high-performing and easily com-prehensible Retrieval-Augmented Generation systems, this work is an important milestone.