Embodied cognitive theories predict that linguistic conceptual representations are grounded and continually represented in real world, sensorimotor experiences. However, there is an on-going debate on whether this also holds for abstract concepts. Grammar is the archetype of abstract knowledge, and therefore constitutes a test case against embodied theories of language representation. Former studies have largely focussed on lexical-level embodied representations. In the present study we take the grounding-by-modality idea a step further by using reaction time (RT) data from the linguistic processing of nominal classifiers in Chinese. We take advantage of an independent body of research, which shows that attention in hand space is biased. Specifically, objects near the hand consistently yield shorter RTs as a function of readiness for action on graspable objects within reaching space, and the same biased attention inhibits attentional disengagement. We predicted that this attention bias would equally apply to the graspable object classifier but not to the big object classifier. Chinese speakers (N = 22) judged grammatical congruency of classifier-noun combinations in two conditions: graspable object classifier and big object classifier. We found that RTs for the graspable object classifier were significantly faster in congruent combinations, and significantly slower in incongruent combinations, than the big object classifier. There was no main effect on grammatical violations, but rather an interaction effect of classifier type. Thus, we demonstrate here grammatical category-specific effects pertaining to the semantic content and by extension the visual and tactile modality of acquisition underlying the acquisition of these categories. We conclude that abstract grammatical categories are subjected to the same mechanisms as general cognitive and neurophysiological processes and may therefore be grounded.