Language acquisition is fundamental for the development of various skills in the early stage of a children’s life. Unfortunately, developmental language disorder (DLD) is the most common developmental disorder during childhood. A common indicator of DLD is that children with such condition struggle to correctly use grammatical forms. Therefore, we focus in this work on automatic grammatical error detection on spontaneous children’s speech. We extend the state of the art by an iterative pseudo labeling scheme to account for the ambiguity of grammatical error labels. Such ambiguity becomes obvious, when it is unclear which word is incorrect, e. g., for agreement errors. In terms of the F1 gain score (FG1) we significantly improve upon the baseline on sentence- and word-level label. On automatic transcriptions of the kidsTALC corpus we increase the sentence-level FG1 from 0.38 to 0.63. Further, our best performing system achieves a recall of 0.45, while maintaining a precision of 0.36.