Diego Collarana
Diego Collarana has a total of 29 co-authors. Here are the top co-authors by number of shared publications:
- Damien Graux
Huawei Research Ltd., Edinburgh, United Kingdom
3 shared publications - Jens Lehmann
Faculty of Computer Science, TUD Dresden University of Technology, Dresden, Germany
2 shared publications - Maria-Esther Vidal
TIB Leibniz Information Centre for Science and Technology
2 shared publications - Mayesha Tasnim
University of Amsterdam
2 shared publications - Afshin Sadeghi
University of Bonn
1 shared publication - Alexander Grass
RWTH Aachen University, Aachen, Germany
1 shared publication
Latest Publications
- Open Access
- 2 September 2026
- Published by IOS Press
- p. 141-159
GraphRAG Best Practices: A Modular Evaluation Framework and Regulatory Case Study
Graph Retrieval-Augmented Generation (GraphRAG) enhances Large Language Models by integrating knowledge graphs into retrieval and generation. Despite growing adoption, no systematic study has isolated the contribution of individual pipeline components to end-to-end performance, leaving practitioners with-out evidence-based configuration guidance. We address this gap with a component-level empirical study spanning three GraphRAG stages, namely graph-based indexing, graph-guided retrieval, and graph-enhanced generation, and identify which design decisions most significantly influence answer quality. Across three datasets (MSMARCO, HotpotQA, and a domain-specific EU banking regulation corpus), we benchmark 17 embedding models, six retriever types with varying traversal depths, nine reranking models, and four graph serialisation formats for prompt construction. Our analysis reveals three key findings: (1) retriever combination strategy and traversal depth are dataset-dependent, with domain-specific corpora favouring shallow traversal and separate reranking; (2) a dual-retriever pairing hypothetical-question embeddings with entity-based retrieval, pruned by a cross-encoder reranker, yields the most robust performance across domains (F1 up to 89.9); and (3) code-like graph serialisation (GraphML) provides the best quality– latency trade-off for generation, likely due to LLM familiarity with structured syntax. Applied to 500 questions over EU banking regulations (CRR + CRD IV), the optimised GraphRAG pipeline achieves win rates of 58–67% against Naive, HyDE, and Hybrid RAG baselines across comprehensiveness, diversity, empowerment, and correctness. All generation components use open-source LLMs (Llama-3.3-70B-Instruct), and datasets are publicly available, supporting reproducibility.
- 2 September 2026
- Published by IOS Press
- p. 141-159
Open AccessGraphRAG Best Practices: A Modular Evaluation Framework and Regulatory Case Study
Graph Retrieval-Augmented Generation (GraphRAG) enhances Large Language Models by integrating knowledge graphs into retrieval and generation. Despite growing adoption, no systematic study has isolated the contribution of individual pipeline components to end-to-end performance, leaving practitioners with-out evidence-based configuration guidance. We address this gap with a component-level empirical study spanning three GraphRAG stages, namely graph-based indexing, graph-guided retrieval, and graph-enhanced generation, and identify which design decisions most significantly influence answer quality. Across three datasets (MSMARCO, HotpotQA, and a domain-specific EU banking regulation corpus), we benchmark 17 embedding models, six retriever types with varying traversal depths, nine reranking models, and four graph serialisation formats for prompt construction. Our analysis reveals three key findings: (1) retriever combination strategy and traversal depth are dataset-dependent, with domain-specific corpora favouring shallow traversal and separate reranking; (2) a dual-retriever pairing hypothetical-question embeddings with entity-based retrieval, pruned by a cross-encoder reranker, yields the most robust performance across domains (F1 up to 89.9); and (3) code-like graph serialisation (GraphML) provides the best quality– latency trade-off for generation, likely due to LLM familiarity with structured syntax. Applied to 500 questions over EU banking regulations (CRR + CRD IV), the optimised GraphRAG pipeline achieves win rates of 58–67% against Naive, HyDE, and Hybrid RAG baselines across comprehensiveness, diversity, empowerment, and correctness. All generation components use open-source LLMs (Llama-3.3-70B-Instruct), and datasets are publicly available, supporting reproducibility.
- Open Access
- 11 September 2023
- Published by IOS Press
- p. 38-52
QALD-9-ES: A Spanish Dataset for Question Answering Systems
Knowledge Graph Question Answering (KGQA) systems enable access to semantic information for any user who can compose a question in natural language. KGQA systems are now a core component of many industrial applications, including chatbots and conversational search applications. Although distinct worldwide cultures speak different languages, the number of languages covered by KGQA systems and its resources is mainly limited to English. To implement KGQA systems worldwide, we need to expand the current KGQA resources to languages other than English. Taking into account the recent popularity that Large-Scale Language Models are receiving, we believe that providing quality resources is key to the development of future pipelines. One of these resources is the datasets used to train and test KGQA systems. Among the few multilingual KGQA datasets available, only one covers Spanish, i.e., QALD-9. We reviewed the Spanish translations in the QALD-9 dataset and confirmed several issues that may affect the KGQA system's quality. Taking this into account, we created new Spanish translations for this dataset and reviewed them manually with the help of native speakers. This dataset provides newly created, high-quality translations for QALD-9; we call this extension QALD-9-ES. We merged these translations into the QALD-9plus dataset, which provides trustworthy native translations for QALD-9 in nine languages, intending to create one complete source of high-quality translations. We compared the new translations with the QALD-9 original ones using languageagnostic quantitative text analysis measures and found improvements in the results of the new translations. Finally, we compared both translations using the GERBIL QA benchmark framework using a KGQA system that supports Spanish. Although the question-answering scores only improved slightly, we believe that improving the quality of the existing translations will result in better KGQA systems and therefore increase the applicability of KGQA w.r.t. the Spanish language domain.
- 11 September 2023
- Published by IOS Press
- p. 38-52
Open AccessQALD-9-ES: A Spanish Dataset for Question Answering Systems
Knowledge Graph Question Answering (KGQA) systems enable access to semantic information for any user who can compose a question in natural language. KGQA systems are now a core component of many industrial applications, including chatbots and conversational search applications. Although distinct worldwide cultures speak different languages, the number of languages covered by KGQA systems and its resources is mainly limited to English. To implement KGQA systems worldwide, we need to expand the current KGQA resources to languages other than English. Taking into account the recent popularity that Large-Scale Language Models are receiving, we believe that providing quality resources is key to the development of future pipelines. One of these resources is the datasets used to train and test KGQA systems. Among the few multilingual KGQA datasets available, only one covers Spanish, i.e., QALD-9. We reviewed the Spanish translations in the QALD-9 dataset and confirmed several issues that may affect the KGQA system's quality. Taking this into account, we created new Spanish translations for this dataset and reviewed them manually with the help of native speakers. This dataset provides newly created, high-quality translations for QALD-9; we call this extension QALD-9-ES. We merged these translations into the QALD-9plus dataset, which provides trustworthy native translations for QALD-9 in nine languages, intending to create one complete source of high-quality translations. We compared the new translations with the QALD-9 original ones using languageagnostic quantitative text analysis measures and found improvements in the results of the new translations. Finally, we compared both translations using the GERBIL QA benchmark framework using a KGQA system that supports Spanish. Although the question-answering scores only improved slightly, we believe that improving the quality of the existing translations will result in better KGQA systems and therefore increase the applicability of KGQA w.r.t. the Spanish language domain.
- 16 June 2023
- Published by Springer Nature
- p. 336-351
Topio: An Open-Source Web Platform for Trading Geospatial Data
- 16 June 2023
- Published by Springer Nature
- p. 336-351
Topio: An Open-Source Web Platform for Trading Geospatial Data
- 25 January 2022
- Published by Elsevier in Knowledge-Based Systems
- Vol. 241
Spatial concept learning and inference on geospatial polygon data
- 25 January 2022
- Published by Elsevier in Knowledge-Based Systems
- Vol. 241
Spatial concept learning and inference on geospatial polygon data
- 10 September 2021
- Published by Springer Nature
- p. 548-564
Embedding Knowledge Graphs Attentive to Positional and Centrality Qualities
- 10 September 2021
- Published by Springer Nature
- p. 548-564
Embedding Knowledge Graphs Attentive to Positional and Centrality Qualities
- 21 July 2021
- Published by Springer Nature
- p. 245-248
A Virtual Knowledge Graph for Enabling Defect Traceability and Customer Service Analytics
- 21 July 2021
- Published by Springer Nature
- p. 245-248
A Virtual Knowledge Graph for Enabling Defect Traceability and Customer Service Analytics
- Open Access
- 16 July 2020
- Published by Springer Nature
- p. 122-146
Chapter 8 Context-Based Entity Matching for Big Data
- 16 July 2020
- Published by Springer Nature
- p. 122-146
Open AccessChapter 8 Context-Based Entity Matching for Big Data
- Open Access
- 13 May 2019
- Published by Association for Computing Machinery (ACM)
- p. 961-965
Summarizing Entity Temporal Evolution in Knowledge Graphs
- 13 May 2019
- Published by Association for Computing Machinery (ACM)
- p. 961-965
Open AccessSummarizing Entity Temporal Evolution in Knowledge Graphs