An approach to unsupervised ontology term tagging of dependency-parsed text using a Self-Organizing Map (SOM)
Seppo Nyrkkö · Digital Humanities in the Nordic and Baltic Countries Publications · 2018
I describe here a machine-learning estimation method for term tagging which can learn semantic disambiguation. The model is trained with a Semantic Web ontology, and a set of sample text documents with a set of concepts tagged, referring to the given ontology. The machine-learning method is based on creating numeric representations, or embeddings, which are based on dependency analysis of the syntactic environment of the word being analyzed. In contrast to many modern neural data-driven models, this model uses a less data-hungry unsupervised clustering method, the Self-Organizing Map (SOM). Based on the observations found with the experimental model, I suggest this can be utilized for populating ontologies with new concepts and terms, and for guessing the best matching ontology concepts for the found terms.