NER/RE strategies for biological graph completion

Natural Language Processing (NLP), a subset of artificial intelligence, facili- tates computer understanding of human language, with techniques like Named Entity Recognition (NER) and Relation Extraction (RE) essential for trans- forming unstructured text into structured, actionable data. In the bio...

Descripción completa

Detalles Bibliográficos
Autor: Caballero Vergés, Biel
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/419447
Acceso en línea:https://hdl.handle.net/2117/419447
Access Level:acceso abierto
Palabra clave:Natural language processing (Computer science)
Processament del Llenguatge Natural
Reconeixement d'Entitats Nombrades
Extracció de Relacions
grafs de coneixement biològic
extracció d'entitats
extracció de relacions
resums de PubMed
ChemProt
BioRED
BioBERT
SciBERT
incrustació de gràfics de coneixement
ComplEx
RotatE
terminologia biomèdica.
Natural Language Processing
Named Entity Recognition
Relation Extraction
biological knowledge graphs
entity extraction
relationship extraction
PubMed abstracts
knowledge graph embedding
biomedical terminology.
Tractament del llenguatge natural (Informàtica)
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Llenguatge natural
Descripción
Sumario:Natural Language Processing (NLP), a subset of artificial intelligence, facili- tates computer understanding of human language, with techniques like Named Entity Recognition (NER) and Relation Extraction (RE) essential for trans- forming unstructured text into structured, actionable data. In the biomedi- cal domain, NER identifies critical biological entities such as genes, proteins, and diseases while RE discovers relationships like gene-disease associations and protein-protein interactions. These methods are particularly valuable for cre- ating biological knowledge graphs, which model complex interactions among biological entities, supporting research in areas like drug discovery and disease mechanisms. However, challenges arise due to domain-specific language and the intricate nature of biological relationships, posing a need for more accurate, scalable extraction methods. This thesis aims to address these challenges by developing NER and RE strate- gies for biological texts, using both rule-based and pretrained models on PubMed abstracts. The study uses biomedical datasets, including ChemProt and BioRED, to evaluate performance in entity and relationship extraction, while focusing on improving precision and scalability for biological knowledge graph comple- tion. Specifically, pretrained models like BioBERT and SciBERT, designed for biomedical NLP, were applied, achieving high accuracy in identifying domain- specific entities and relationships. Additionally, these extracted entities and relationships were integrated into a biological knowledge graph, utilizing knowl- edge graph embedding techniques, such as ComplEx and RotatE, to represent the complex associations within biological systems. Results indicate that pretrained models excel in capturing biomedical terminol- ogy, with BioBERT showing particular strength in disease-related entity extrac- tion. Integration of text-mined data into knowledge graph embeddings enhances accuracy in relationship predictions, emphasizing the potential of combining structured and unstructured data for dynamic graph completion. By improv- ing both NER and RE techniques, this research contributes new methodologies for automating biological knowledge graph construction, supporting future ad- vancements in fields like genomics, drug discovery, and systems biology.