Action Type induction from multilingual lexical features

摘要: This paper presents a vector representation and clustering of action concepts based on lexical features extracted from IMAGACT, multilingual multimodal ontology actions in which are represented through video prototypes. We computed vectors for 1,010 concepts, where the dimensions correspond to verbs 10 languages. Finally, an unsupervised method has been applied these data order discover classes typological closeness. Those clusters not language-specific or language-biased, thus constitute inter-linguistic classification domain.