To examine the processing of phonological and configurational information in word recognition in discourse reading, we conducted two experiments using the self-paced reading paradigm. The materials were three-sentence discourses, in each of which the last word of the second sentence and the third word from the end of the last sentence formed a prime--target pair. The discourse in which the target word (T) was semantically congruent or incongruent with the prime word was converted into a new version by replacing the T with its homophone or with the control word (con-T) in Experiment 1. Similarly, the Ts were replaced by words that were similar to them in configuration or by the con-Ts in Experiment 2. We adopted mixed-effects modeling to analyze the participants' reading times to the targets, the first words after the targets, and the second words after the targets. It is concluded that the processing of phonological information begins earlier than that of configurational information in activating the semantic representations for the upcoming words that fit the context in discourse reading.