The authors have recently constructed an experimental natural language understander named IMAGES-I (I__-nterlingual understanding M__-odel A__-iming at ___-neral purpose S__-ystem-I). This system consists of a syntactic and a semantic processor which can work independently of each other unlike most of the conventional systems. The syntactic processor transforms surface structures into surface dependency structures and sends them to the semantic processor. This paper discusses especially the semantic processing of IMAGES-I. IMAGES-I is equipped with a dictionary of individual word meanings (SED). Each word meaning consists of a pair of "concept part" and "connection part". The semantic processor of IMAGES-I translates surface dependency structures into semantic structures which are combined word-concepts (i. e. concept parts) and gradually constructed according to the connection parts of the words in some governor-dependent relation. If the processor fails in generating any semantic structure, it rejects the corresponding depedency structure. On the other hand, successfully generated semantic structures are analyzed according to SED, and semantic anomalies, ambiguities and paraphrase relations are detected.