Three experiments examined whether the process of lexical inferences differs according to the direction of contextual elaboration using a semantic relatedness judgment task. In Experiment 1, Japanese university students read English sentences where target unknown words were semantically elaborated by prior contextual information (forward lexical inferences) and by subsequent contextual information (backward lexical inferences). Accuracy of semantic relatedness judgments, and reaction times to target-probe words demonstrated that participants could activate lexical inferences in either case. These findings were replicated in Experiments 2 and 3, where sentence chunks were presented one at a time to prevent the use of a rereading strategy. Taken together, the results suggest that whereas the processes of forward lexical inferences are involved in context-word integration by anticipating upcoming word meanings, those of backward lexical inferences are operating an unknown word in working memory, finding highly elaborative contexts, and then integrating the contextual message with the possible meaning.