Applying Weighted Graph Embeddings to Turkish Metaphor Detection

Metaphor is a common literary mechanism that allows abstract concepts to be conceptualised using more concrete terminology. Existing methods rely on either end-to-end models or hand-crafted pre-processing steps. Generating well-defined training datasets for supervised models is a time-consuming operation for this type of problem. There is also a lack of pre-processing steps for resource-poor natural languages. In this study, we propose an approach for detecting Turkish metaphorical concepts. Initially, we collect non-literal concepts including their meaning and reference sentences by employing a Turkish dictionary. Secondly, we generate a graph by discovering super-sense relations between sample texts including target metaphorical expressions in Turkish WordNet. We also compute weights for relations based on the path closeness and word occurrences. Finally, we classify the texts by leveraging a weighted graph embedding model. The evaluation setup indicates that the proposed approach reaches the best F1 and Gmean scores of 0.83 and 0.68 for the generated test sets when we use feature vector representations of the Node2Vec model as the input of the logistic regression for detecting metaphors in Turkish texts.