Since Harnad (1990) pointed out the symbol grounding problem, cognitive science research has demonstrated that grounding in perceptual or sensorimotor experience is crucial to language. Recent embodied cognition theories have argued that language is more important for grounding abstract than concrete words; abstract words are grounded via language. Distributional semantics has recently addressed the embodied nature of language and proposed multimodal semantic models. However, these models are not cognitively plausible because they do not address the recent embodiment view of abstract concepts. Therefore, we propose a novel multimodal distributional semantics in which abstract words are represented indirectly through grounded representations of their semantically related concrete words. A simulation experiment demonstrated that the proposed model achieved better performance in computing the word similarity than other multimodal or text-based distributional models. This finding suggests that the indirect embodiment view is plausible and contributes to the improvement of multimodal distributional semantics.