Action Type induction from multilingual lexical features

Abstract

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