What would be best practice in indicating the source of text samples used to illustrate some concept or analysis result in linguistics or computational linguistics?
For example, I might want to use the sentence
If you had told someone in 2012 that in just two years hence, the eurozone would remain bonded together but the United Kingdom might not, they would have thought you insane.
from a recent The New York Times article (http://www.nytimes.com/2014/09/10/upshot/why-does-scotland-want-independence-its-culture-vs-economics.html?abt=0002&abg=0), augmented with some annotation, in order to illustrate how different anaphora resolution methods would work or fail for this real life example. Copyright fair use policies allow to use such limited extracts without any permission, and using them is common practice, but the attribution and quoting seems unclear.
How, and if, should I quote the source of that sentence within a paper?
Would publication style (say, short paper vs. a dissertation) affect it?
Would density of examples (2 examples in whole paper vs 20 different examples in a single page) matter?
Would indirect sources change it? E.g. if the sentences are taken from, say, British National Corpus - but, naturally, they originally come from some different publication.
Language utterance sources are different from other references
At least in computational linguistics, general practice clearly is to not include the source of language utterance or sentence examples together with normal references. There are various approaches seen in practice among respectable publications: no referencing at all, a footnote reference for the sentence, a mention in article text, a single reference to a whole corpus (by referencing not the source but some authoritative paper about the corpus), etc. I'm wondering about those options, which would be preferable and which should be avoided.
Avoiding the whole problem by using sentences made up by myself is not a good solution - artificial examples are biased towards using constructions that behave in the same way as you and your tools expect, so real life examples of natural language from domain-representative sources are strictly preferred.