Most object-based attention studies use objects defined bottom-up by Gestalt principles. In the present study, we defined objects top-down, using Chinese words that were seen as objects by skilled readers of Chinese. Using a spatial cuing paradigm, we found that a target character was detected faster if it was in the same word as the cued character than if it was in a different word. Because there were no bottom-up factors that distinguished the words, these results showed that objects defined by subjects' knowledge--in this case, lexical information--can also constrain the deployment of attention.