The present study investigated the way in which the activation of semantic representations at the morpheme level affects the processing of two-kanji (morpheme) compound words. Three types of Japanese two-kanji compound words were used as stimulus items: (1) words consisting of two kanji representing opposite concepts (e.g., long + short = length), (2) words consisting of two kanji representing similar concepts (e.g., soft + flexible = pliable), and (3) control words consisting of two closely bound kanji (e.g., wild + field = wilderness). Words consisting of kanji of opposite concepts (M = 768 ms for LD and M = 645 ms for naming) were processed more slowly for lexical decision, but not for naming, than words with kanji of similar concepts (M = 743 ms for LD and M = 636 ms for naming), both of which were, furthermore, processed more slowly for lexical decision and naming than the control compound words (M = 716 ms for LD and M = 590 ms for naming). These results were explained in the framework of the multiple-level interactive-activation model as follows. Because kanji morphemes of opposite and similar concepts are semantically activated both as morpheme units and compound word units, semantic representations of the two morphemes and the compound word which they create compete with each other at the concept level, which slows down lexical decision and naming of the compound word.