علي فتي

In recent years, the rapid growth of unstructured textual data has highlighted the need for efficient methods to extract, organize, an‎d represent knowledge. Ontology, as a core component of knowledge engineering, plays a crucial role in enhancing natural language processing systems by modeling concepts an‎d their semantic relationships. However, relation extraction, as a key step in automatic ontology construction, faces several challenges, including semantic complexity, strong contextual dependency, an‎d limited capability for discovering novel knowledge—particularly in the domain of pharmacology. In this thesis, we propose a two-stage framework based on transfer learning, large language models, an‎d graph-based approaches to improve the extraction an‎d enrichment of drug-related relations. In the first stage, the DREaM method is employed to extract an‎d classify relationships among pharmaceutical entities from scientific texts in the PubMed database. By transferring knowledge from the ACORD framework an‎d leveraging transfer learning, this method achieves an average precision of 73.5%, recall of 72.2%, an‎d an F1-score of 71.2%, demonstrating more stable an‎d reliable performance compared to direct extraction using large language models on specialized biomedical data. In the second stage, the PARADOX framework is introduced to enrich the resulting ontology graph by identifying contradictions among drug-related relations using Siamese neural networks. Ultimately, from the can‎didate contradictions extracted from PubMed texts, 39 valid drug-related contradictions are identified an‎d represented in the form of a contradiction graph, leading to increased semantic coverage an‎d improved quality of the drug knowledge graph. Overall, the results indicate that integrating transfer learning, large language models, an‎d graph-based approaches can effectively enhance relation extraction accuracy an‎d facilitate the construction of richer an‎d more interpretable ontologies in the pharmacological domain.

اﺳﺘﺨﺮاج راﺑﻄه , پردازش زبان طبيعي , يادگيري ماشين , ساخت هستان‌شناسي , گراف دانش , تناقض‌يابي

Relation Extraction , Natural Language Processing , Machine Learning , Ontology Construction , Knowledge Graph , Contradiction Detection

Ali Fata

Dr. Behrouz Minaei-Bidgoli