Evaluating data imbalance approaches for classifying semantic relations using machine learning and word embeddings

• Autor(a) principal: Domingues, Gabriel Couto
• Data de Publicação: 2023
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: eng
• Título da fonte: Repositório Institucional da UFRGS
• Texto Completo: http://hdl.handle.net/10183/267624
• topic: Aprendizado de máquina; Redes neurais; Semântica computacional; Word Embeddings; Supervised Learning; Ontologies; Knowledge Graphs; WordNet

Explicit knowledge models are artifacts that represent domain knowledge in an explicit way and can be used in different ways, including structuring data, supporting information retrieval and reasoning. The identification and classification of semantic relationships between concepts is a critical task in the development of knowledge models. This work investigates the use of machine learning approaches and pre-trained static word embeddings to classify semantic relationships between concepts, evaluating different techniques to deal with the challenges imposed by data imbalance in this context. We proposed a methodology for building datasets for the task of semantic relationship classification from word embeddings using WordNet as a semantic reference. By applying the proposed methodology, we generated two different datasets, with two variations, for the target task. Finally, we evaluated a set of general approaches for dealing with data imbalance in classification tasks. Our results indicated that while some strategies like SMOTE showed promise in specific metrics, the baseline model consistently achieved superior performance in terms of F1 score.