What are the most authoratative references on the notion of shallow parsing (also known as text chunking) in NLP? I am looking for definitions, examples, and discussion of strings of words that qualify as chunks in the sense of shallow parsing. Some surfing on the internet has revealed that the followng article was seminal in this area:

Abney, Steven (1991), Parsing By Chunks. Kluwer Academic Publishers, pp. 257–278.

That article is, however, quite dated. Are there prominent articles or book chapters that explore the concept of text chunking in a broad sense?


Maybe this link will be useful for you: NP Chunking (State of the art).

  • 1
    Additionally, the references on this page point to a lot of the classic texts (in particular Church 88), and some newer results as well: ifarm.nl/erikt/research/np-chunking.html – Fred Jun 30 '15 at 20:30
  • Thanks for the link. I've taken a quick look. There does indeed appear to be a paper or two there that provide some guidance about how one identifies chunks syntactically. Note, however, that most of the papers and discussions are concerned with NP chunks. The status of NPs as chunks strikes me as uncontroversial and thus uninteresting. – Tim Osborne Jul 3 '15 at 5:50

I never heard of shallow parsing, but I have a theory that sort of does it. I don't often see a question that seems to license me talking about it.

Consider a variant of context free grammar in which I. heads are not assigned to grammatical types (N, A, V, P, ...) but are rather introduced directly by phrase structure rules (S1 -> NP1 loves NP2), and II. grammatical types are distinguished by their degree of embedding, which I note with a digit suffix (0, 1, 2, 3). My preceding illustration means that a finite clause (S1) consists of a nominative NP (NP1) followed by pronunciation "loves" followed by an accusative NP (NP2).

The sense in which this gives a shallow parsing is that much of the embeddedness which requires huge tall trees in conventional syntactic theory is there in the psr. Much of the apparent branching structure accompanies lexical forms of verbs (which are psrs). I assign intrinsic embeddededness levels to the various auxiliaries of English and the various adverb types (Adv0 is performative, Adv1 is sentential, Adv2 is manner, Adv3 is degree).

You can make tall trees if you want, though. Corresponding to the principle of the transformational cycle, psrs are expanded always starting with the most embedded parts.

So if you think of parsing as associating psrs with strings of words and trees, there are many fewer associations required for a tree of a given depth, so the parse is shallow in that sense.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.