Speech dialog systems need to deal with various kinds of ill-formed speech inputs that appear in natural human-human dialog. Self-correction (or speech-repair) is a particularly problematic phenomenon. Although many ways of dealing with selfcorrection have been proposed, these have limitations in both detecting and correcting for this phenomenon. In this paper, we propose a method to overcome these problems in Japanese speech dialog. We evaluate the proposed method using our speech dialog corpus and discuss its limitations and the work that remains to be done.