This paper introduces a semi-supervised method for automatic speech act recognition in emails and forums, addressing the challenge of limited labeled data in these genres. The method utilizes labeled data from the SwitchboardDAMSL and Meeting Recorder Dialog Act databases and applies domain adaptation techniques to a large volume of unlabeled email and forum data. It employs automatically extracted features, such as phrases and dependency trees (subtree features), for semi-supervised learning. Empirical results show that this approach effectively improves speech act recognition in email and forum contexts.