Deep natural language processing tools can provide a lot of information about the texts they process, but because they lack robustness they are often not used in natural language applications where that information would be useful. The ongoing theme of my research has been to investigate methods of using deep NLP techniques robustly, through integration with less informative shallow processing techniques. In this talk I will give an overview of the research I did for my Masters thesis, where I used shallow NLP as a fallback when deep processing failed in a Japanese question answering system. I will then outline my current PhD proposal, looking at methods of integrating deep and shallow methods, rather than naively falling back to the shallow tools. Feedback and suggestions would be very welcome.