Before anything, try to preserve the info about "boundaries" which comes in the input text.
(if such info has not readily be lost, your question implies that maybe the tokenization has readily been done)
During the tokenization (word parsing, in this case) process, look for patterns that may define expression boundaries (such as punctuation, particularly periods, and also multiple LF/CR separation, use these. Also words like "the", can often be used as boundaries. Such expression boundaries are typically "negative", in a sense that they separate two token instances which are sure to not be included in the same expression. A few positive boundaries are quotes, particularly double quotes. This type of info may be useful to filter-out some of the n-grams (see next paragraph). Also word sequencces such as "for example" or "in lieu of" or "need to" can be used as expression boundaries as well (but using such info is edging on using "priors" which I discuss later).
Without using external data (other than the input text), you can have a relative success with this by running statistics on the text's digrams and trigrams (sequence of 2 and 3 consecutive words). Then [most] the sequences with a significant (*) number of instances will likely be the type of "expression/phrases" you are looking for.
This somewhat crude method will yield a few false positive, but on the whole may be workable. Having filtered the n-grams known to cross "boundaries" as hinted in the first paragraph, may help significantly because in natural languages sentence ending and sentence starts tend to draw from a limited subset of the message space and hence produce combinations of token that may appear to be statistically well represented, but which are typically not semantically related.
Better methods (possibly more expensive, processing-wise, and design/investment-wise), will make the use of extra "priors" relevant to the domain and/or national languages of the input text.
- POS (Part-Of-Speech) tagging is quite useful, in several ways (provides additional, more objective expression boundaries, and also "noise" words classes, for example all articles, even when used in the context of entities are typically of little in tag clouds such that the OP wants to produce.
- Dictionaries, lexicons and the like can be quite useful too. In particular, these which identify "entities" (aka instances in WordNet lingo) and their alternative forms. Entities are very important for tag clouds (though they are not the only class of words found in them), and by identifying them, it is also possible to normalize them (the many different expressions which can be used for say,"Senator T. Kennedy"), hence eliminate duplicates, but also increase the frequency of the underlying entities.
- if the corpus is structured as a document collection, it may be useful to use various tricks related to the TF (Term Frequency) and IDF (Inverse Document Frequency)
[Sorry, gotta go, for now (plus would like more detail from your specific goals etc.). I'll try and provide more detail and pointes later]
[BTW, I want to plug here Jonathan Feinberg and Dervin Thunk responses from this post, as they provide excellent pointers, in terms of methods and tools for the kind of task at hand. In particular, NTLK and Python-at-large provide an excellent framework for experimenting]
