Revealing the relations between pitch accent types and the informational status of words requires a refined discourse analysis of spontaneous speech. A cooperative unscripted task in which subjects gave instructions for decorating Christmas trees successfully induced production of target adjective-noun pairs conveying new/given and contrastive information. Adapting Grosz and Sidner’s intention-based discourse analysis [1], each target word was tagged for its newness or givenness and also for contrastiveness at both the discourse level and the discourse segment level. The analyses show that contrastiveness was a good predictor of accent type (L+H*), and that the finer-grained discourse segment level analysis was somewhat better than the discourse level in predicting the presence or absence of accent. Local word position (adjective or noun) interacted with both contrastiveness and discourse segmentation in the assignment of accent.