This paper presents a brief description of Explanation-Based Learning (EBL), and argues that it is an approach to machine learning with significant potential for use in discourse processing. More specifically, EBL can be used by systems that model discourse generation as goaldriven behavior, and that model discourse interpretation as recognizing the speaker’s discourse goals. As evidence, we describe the discourse architecture of the NL-Soar dialogue system. The architecture uses the EBL capabilities of Soar to compile discourse recipes during discourse planning. The recipes can be reused to speed up discourse generation and as a knowledge source during discourse interpretation.