The only answer I can think about is: it depends on the questions you want to answer ! Sorry but I seriously cannot do better. One could argue the deeper the better... although most systems, will focus on factoid questions, where shallow parsing is normally sufficient. Thus, in EAGLi (http://eagl.unige.ch/eagli/), we use a deep syntactic parser (GB) but in operation we only use a marginal fraction of the power of the parser because we intensively use onto-terminological resources. So, for such a QA system, a NP recognizer should be sufficient.

Thank you.. I like to extract patterns from the text documents. For eg. hiv cause of aids, with a pattern "cause of" expresses causation relation. poem consist of stanza with a pattern "consist of" expresses part-whole relation.. Will shallow parsing sufficient? Also is coreference resolution and textual entailment possible with shallow parsing..

For all the examples of patterns you give, I think shallow parsing or even POS tagging provided that you write the patterns for the extraction of NPs will be sufficient. For anaphora that is a totally different story. Here you have issues such as (very) long distance dependencies, combined with case & semantic inference constraints. Neither shallow nor deep parsing will be sufficient here. Most syntactic parsers and theories are applied at the level of the sentence. To go beyond sentence boundaries you need a grammar of the text. This is discourse analysis (argumentative structure, rhetorical structure...), not syntactic parsing. Your project seems ambitious, I would recommend to prioritize your objectives :-)

Thank you..