We have been developing a neural natural language understanding system oriented towards the extraction of information relevant to a task. We have implemented a version of this system for an air travel reservation task. We have studied systems that map input sequences to output sequences, and conceived a recurrent neural system for conceptual decoding. Robustness, learnability and flexible integration of different information sources are some of the attractive features of our model. Training Neural Networks to capture long-term dependencies remains an important challenge. We present results of our attempts to capture global constraints between noncontiguous semantic patterns. We propose a new system that combines our connectionist model and an N-Best like module for higher-level post segmental treatment.