Abstract
In this thesis, we propose a methodology to guide the process of designing latent structures for a specific structure prediction task. Guided by such a methodology, we demonstrate designing latent structures as explicit structure in the output space for two tasks – Target Sentiment Analysis and Chinese Address Parsing, respectively. Through extensive experiments and detailed analysis, the results show that approaches with such designed latent structures in the output space outperform their competitive baselines significantly. Furthermore, we also demonstrate how to design latent structures as implicit structure in the input space. We argue that the latent structures designed in both the output space and the input space are crucial for building a more successful model for structure prediction tasks through extensive experiments.