In natural languages the default specification of arguments of verbs is often omitted in the surface form. The value of these arguments can be filled by lexical knowledge or commonsense knowledge of human readers, but it is a difficult task for computer programs. Here, we address the need for commonsense knowledge in computational lexicons, and discuss the requisite lexical knowledge of computational lexicons in the language-to-vision application CONFUCIUS. The underspecification problem in natural language visualisation is examined. We compare existing computational lexicons such as WordNet, FrameNet, LCS database, and VerbNet, and show how lexical knowledge in a generative lexicon can be used for disambiguation and commonsense inferencing to fill unspecified argument structures for the task of language visualisation. The possibility of lexical inference with WordNet is explored in order to extract default and shadow arguments of verbs, and in particular, the default argument of implicit instruments/themes of action verbs, which can be used to improve CONFUCIUS' automated language-to-vision conversion through semantic understanding of the text, and to make animation generation more robust by employing the commonsense knowledge included in (or inferred from) lexical entries.