A landmark work on this topic is AbstFinder. Several other strategies exist. Original documents from the Universe of Discourse are information sources to be used in eliciting requirements. They are of fundamental importance.
However, we must be careful when and how to use text-mining. Way back a simple text-mining strategy was commonly used by some OO practicioners: underlining words in a given text, usually a requirements document in natural language written by clients.
Such strategy is not the best choice.
See what Mitchell Lubars, Colin Potts, and Charles Richter wrote in a 1993 ICSE paper.
"Some protagonists of OOA advocate a bottom-up
strategy in which the analys tunderlines or highlights all
the noun phrases in the source material[Rum91].
This produces a list of candidate objects that must then
be pruned according to certain guidelines. This strategy
is sensible for small problems: objects are likely to be
referred to by noun phrases; making the list requires
little judgement and is almost trivial: and object-oriented
methods make many useful recommendations to help the
analyst prune the list.
However,these tasks become overwhelming for problems
of the size we faced. Listing the noun phrases in a
500-page requirements document is a daunting task
of questionable value. A long, unorganized list is
not a good starting point for the next stages of analysis."