In the first case, the system is not actually doing deep understanding, and in the second case the system is not robust as it is restricted to a domain-specific lexicon, grammar, ontology, etc. We first propose to develop a formalism, which allows one to represent both shallow and deep semantic analysis under a unified framework as the first step toward robustness. We then suggest using a deep understanding module as a core and leveraging some external resources to broaden the coverage of the system. External resources such as existing world knowledge resources, unlabeled data, etc. help to learn new information, which enables the system to understand new concepts and relations.