Neural Entity Linking for Question Answering over Knowledge Graphs / Daniel Vollmers ; 1. Reviewer Prof. Dr. Axel-Cyrille Ngonga Ngomo (Department of Computer Science Paderborn University), 2. Reviewer Prof. Dr. Ricardo Usbeck (Institute of Information Systems Leuphana University Lüneburg) ; Supervisors Prof. Dr. Axel-Cyrille Ngonga Ngomo, Dr. Hamada Mohamed Abdelsamee Zahera. Paderborn, 2026
Inhalt
- Titlepage
- Abstract
- Acknowledgement
- Declaration
- Contents
- 1 Introduction
- 2 Background
- 2.1 Basic Natural Language Processing Tasks
- 2.2 Language Models and Neural Networks
- 2.2.1 Neural Networks
- 2.2.2 Recurrent Neural Networks
- 2.2.3 Encoder-Decoder Architecture
- 2.2.4 Large Language Models
- 2.2.5 LLM Implementations
- 2.3 Knowledge Graphs and Querying
- 2.4 Core Tasks in Knowledge Graph Question Answering
- 2.4.1 Information Retrieval
- 2.4.2 Entity Linking and Keyphrase Extraction
- 2.4.3 Semantic Parsing
- 2.4.4 Knowledge Extraction for Semantic Parsing
- 2.5 Evaluation Metrics
- 3 State of the Art
- 3.1 Knowledge Extraction
- 3.1.1 Information Retrieval and Noise-driven Optimization
- 3.1.2 Entity Linking
- 3.1.3 Keyphrase Extraction
- 3.2 Knowledge Graph Question Answering
- 4 Contextual Augmentation for Entity Linking using Large Language Models
- 4.1 Overview
- 4.2 Approach
- 4.3 Ablations
- 4.4 Experiments
- 4.4.1 Experimental Setup
- 4.4.2 Evaluation
- 4.4.3 Datasets
- 4.4.4 Comparison to Baseline Approaches (S1)
- 4.4.5 Evaluation of Foundational Models (S2)
- 4.4.6 Evaluation of Augmentation Strategies (S3)
- 4.4.7 Comparison of Different LLM Models
- 4.5 Limitations
- 4.6 Conclusion
- 5 Evaluating Noisy Optimization in Fine-tuning LMs for Neural Ranking
- 5.1 Overview
- 5.2 Methodology
- 5.3 Experiments
- 5.3.1 Datasets
- 5.3.2 Training and Evaluation Setup
- 5.3.3 Analysis of the Effectiveness of Noise Injection (S1)
- 5.3.4 Comparison of Noise Injection Strategies (S2)
- 5.3.5 Influence of Noise Injection on Convergence (S3)
- 5.4 Optimal Noise Injection per Model and Dataset
- 5.5 Conclusion
- 6 Keyphrase Extraction using Language Models and Knowledge Graphs
- 6.1 Overview
- 6.2 Approach
- 6.2.1 Problem Formulation
- 6.2.2 Present Keyphrase Extraction (PKE)
- 6.2.3 Keyphrase Linking and Absent Keyphrase Generation (AKG)
- 6.2.4 Keyphrases Semantic Matching
- 6.3 Experiments
- 6.4 Conclusion
- 7 UniQ-Gen: Unified Query Generation across Multiple Knowledge Graphs
- 7.1 Overview
- 7.2 Approach
- 7.3 Experiments
- 7.3.1 Datasets
- 7.3.2 Experiment Setup
- 7.3.3 Evaluation
- 7.3.4 Comparison of Unified Model and Single KG models (S1)
- 7.3.5 Comparison with Baseline Models (S2)
- 7.3.6 Influence of KG Knowledge on the Model Performance (S3)
- 7.3.7 Experiments with Different Training Data (S4)
- 7.4 Conclusion
- 8 Knowledge Graph Question Answering using Graph-Pattern Isomorphism
- 8.1 Overview
- 8.2 Approach
- 8.2.1 Question Preprocessing
- 8.2.2 Graph-Isomorphism Detection and Template Classification
- 8.2.3 Information Extraction
- 8.2.4 Query Building
- 8.2.5 Ranking
- 8.3 Experiments
- 8.3.1 Datasets
- 8.3.2 Classification Evaluation (S1)
- 8.3.3 Baseline Experiments (S2)
- 8.3.4 Ablation Study (S3)
- 8.4 Conclusion
- 9 Evaluation of ERL for Question Answering over Knowledge Graphs
- 9.1 Overview
- 9.2 Approaches and Pipelines
- 9.3 Query Generation
- 9.4 Experiments
- 9.4.1 Extraction of KG Knowledge for Questions (RQ3)
- 9.4.2 Influence of ERL Frameworks on Query Prediction (RQ6)
- 9.5 Conclusion
- 10 Application: Enhancing Answers Verbalization using Large Language Models
- 11 Conclusion and Future Work
- 11.1 Summary
- 11.1.1 Contextual Augmentation for Entity Linking and Question Answering
- 11.1.2 Noisy Optimization in Fine-tuning LMs for Neural Ranking
- 11.1.3 Keyphrase Extraction using Language Models and Knowledge Graphs
- 11.1.4 UniQ-Gen: Unified Query Generation across Multiple Knowledge Graphs
- 11.1.5 Knowledge Graph Question Answering using Graph-Pattern Isomorphism
- 11.1.6 Evaluation of ERL for Question Answering over Knowledge Graphs
- 11.1.7 Application: Enhancing Answers Verbalization using Large Language Models
- 11.2 Future Work
- Bibliography