Text-Graph Encoders and Retrieval-Augmented Generation
| Type of publication: | Thesis |
| Citation: | Coman_THESIS_2025 |
| Year: | 2025 |
| Month: | December |
| School: | EPFL |
| URL: | https://infoscience.epfl.ch/en... |
| Abstract: | Modern information-seeking systems increasingly rely on Large Language Models (LLMs) paired with external knowledge sources to generate accurate, context-aware responses. However, current Retrieval-Augmented Generation (RAG) pipelines face limitations in conversational question answering, including limited retrieval coverage and high reader (generator) latency, which affect answer quality and system efficiency. Evaluating responses grounded in retrieved evidence also presents challenges. Reward Models (RMs), typically trained on general preference data, are not equipped to distinguish between plausible responses based on parametric knowledge alone and those grounded in non-parametric evidence retrieved at inference time, as their training lacks retrieval-aware supervision and evidence-conditioned evaluation signals. Although the shift from parametric to non-parametric knowledge has enabled models to access more up-to-date and domain-specific information, most systems continue to rely solely on unstructured (text) inputs. This unstructured evidence often contains implicit, incomplete, or scattered information, making it difficult for models to piece together accurate responses. A promising direction is to incorporate structured non-parametric knowledge, such as knowledge graphs, which make relational information explicit and easier to leverage during inference. This requires architectures capable of jointly encoding unstructured and structured sources in a unified and effective manner. This thesis explores methods to enhance RAG and text-graph encoders through two complementary lines of investigation. The first part focuses on improving the effectiveness and efficiency of retrieverâ reader pipelines, by introducing a lightweight reranking component between the retriever and reader, coupled with targeted fine-tuning steps that increase retrieval coverage and enable the reader to operate on fewer but more relevant passages. Additionally, this part proposes a methodology to adapt reward models for evaluating responses in the RAG setting. By repurposing existing question answering datasets into contextual preference pairs that reflect RAG-specific criteria, it enables the training of evaluation models that prioritise grounded, evidence-based answers over those relying solely on parametric knowledge. The second part investigates how attention-based models can be extended to incorporate explicit relational structure. It introduces a mechanism for jointly encoding unstructured (text) and structured (graph) information within a unified attention framework, enabling the model to draw on both types of input when building representations. This facilitates targeted and context-aware aggregation of information across multiple spans of text. This part further explores how these models can operate over knowledge graphs in inductive settings, where entities and relations encountered at inference time have not been seen during training. It shows that incorporating explicit relational information can reduce the reliance on text encoders to predict links between entities from unstructured input alone. This thesis aims to show the potential of bringing together retrieval-based architectures and structured knowledge representations for building LLM-based information-seeking systems that can ground their generation in retrieved evidence, whether unstructured or structured, and access both types of information seamlessly via the attention mechanism. |
| Keywords: | attention, conversational question answering, Graph, Link Prediction, reader, reranker, Retrieval-augmented generation, retriever, reward, structured, text, unstructured |
| Authors: | |
| Added by: | [UNK] |
| Total mark: | 0 |
|
Attachments
|
|
|
|
|
Notes
|
|
|
|
|
