A carregar...
Publicação
Compound matching for multiple ontologies
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.04 MB | Adobe PDF |
Orientador(es)
Resumo(s)
As áreas científicas multidisciplinares como a Biomédica, usam normalmente redes de ontologias para suportar aplicações como anotação, integração, pesquisa e análise de dados. Estas redes podem ser construídas usando técnicas de correspondência de ontologias, no entanto a maioria das abordagens existentes é limitada a correspondências entre duas ontologias, sendo a grande maioria das equivalências simples. Em cenários de múltiplos domínios, é necessário encontrar correspondências mais complexas, que podem envolver várias ontologias, ou seja, correspondências compostas. Esta dissertação propõe um novo algoritmo de alinhamentos compostos, capaz de criar correspondências entre uma classe de origem e uma expressão de classe, relacionando múltiplas classes de múltiplas ontologias alvo. Trata das limitações de abordagens anteriores, que apenas consideraram duas classes de duas ontologias alvo. O algoritmo é baseado nas abordagens eficientes de correspondência léxica do AgreementMakerLight. Uma avaliação automática foi realizada contra alinhamentos de referência parciais usando métricas de avaliação clássicas e também novas, mais adequadas para a avaliação do alinhamento composto. Apesar dos resultados com métricas clássicas serem algo limitados (um facto ao qual não ajuda a incompletude dos alinhamentos de referência), as novas métricas de avaliação, projetadas para medir a utilidade de uma correspondência num cenário de alinhamento interativo, são promissoras, com menor precisão, mas com valores de recall entre 80-98%.
Multi-domain areas, such as the biomedical field, routinely employ networks of ontologies to support applications such as data annotation, integration, search and analysis. These networks can be built using ontology matching techniques, however most existing approaches are limited to matches between two ontologies, the large majority being simple equivalences. In multi-domain scenarios, there is a need to discover more complex mappings, that may involve multiple ontologies, i.e. compound mappings. This thesis proposes a novel compound matching algorithm, able to compose mappings between a source class and a class expression relating multiple classes from multiple target ontologies. It addresses the limitations of previous approaches that only considered two target classes from two target ontologies. The algorithm is based on the efficient lexical matching approaches in AgreementMakerLight. An automatic evaluation was carried against partial reference alignments using both classical and novel evaluation metrics more suited to compound alignment evaluation. Despite results with classical metrics being rather poor (a fact not helped by the incompleteness of the reference alignments), the novel evaluation metrics, designed to measure the usefulness of a mapping in an interactive alignment scenario are promising, with lower precision, but recall values in the 80-98% range.
Multi-domain areas, such as the biomedical field, routinely employ networks of ontologies to support applications such as data annotation, integration, search and analysis. These networks can be built using ontology matching techniques, however most existing approaches are limited to matches between two ontologies, the large majority being simple equivalences. In multi-domain scenarios, there is a need to discover more complex mappings, that may involve multiple ontologies, i.e. compound mappings. This thesis proposes a novel compound matching algorithm, able to compose mappings between a source class and a class expression relating multiple classes from multiple target ontologies. It addresses the limitations of previous approaches that only considered two target classes from two target ontologies. The algorithm is based on the efficient lexical matching approaches in AgreementMakerLight. An automatic evaluation was carried against partial reference alignments using both classical and novel evaluation metrics more suited to compound alignment evaluation. Despite results with classical metrics being rather poor (a fact not helped by the incompleteness of the reference alignments), the novel evaluation metrics, designed to measure the usefulness of a mapping in an interactive alignment scenario are promising, with lower precision, but recall values in the 80-98% range.
Descrição
Tese de mestrado, Bioinformática e Biologia Computacional (Bioinformática) Universidade de Lisboa, Faculdade de Ciências, 2019
Palavras-chave
Alinhamento de Ontologias Alinhamento Complexo de Ontologias Alinhamento Composto de Ontologias Ontologias Biomédicas Teses de mestrado - 2019