We are getting better and better in solving various subproblems in Natural Language Processing (NLP), such as parsing, coreference or relation extraction; however, once assembled into an end-to-end system of the traditional pipeline architecture, errors cascade and magnify. The principle goal of this project is to enable new generation of NLP applications in which information flow is bidirectional, and acquired downstream knowledge increases the robustness of upstream processing. Specifically, we want to investigate bidirectional flow in scenarios where downstream processing can acquire knowledge in very rich representations, and learn from massive amounts of unlabeled data. While this goal is motivated by the need for more accurate NLP, it also relates to the fundamental problem building artificial cognitive systems that adapt to their environment, seamlessly connect complex layers of abstraction and never stop learning. The work will have direct applications, for example, in extracting meta-data from media archives, biomedical text mining and information extraction from clinical texts