Abstract
We describe a Chinese temporal annotation experiment that produced a sizable data set for the TempEval-2 evaluation campaign. We show that while we have achieved high inter-annotator agreement for simpler tasks such as identification of events and time expressions, temporal relation annotation proves to be much more challenging. We show that in order to improve the inter-annotator agreement it is important to strategically select the annotation targets, and the selection of annotation targets should be subject to syntactic, semantic and discourse constraints.