I was taught X' theory in generative grammar lectures, only to find out that NLP uses mainly simpler grammars, such as HPSG or statistical models such as hidden Markov models.

Is X' of any use for NLP at its current state of the art?

  • 2
    Yes, it's called information theory. ;) – Alan H. Sep 13 '11 at 22:54
  • 1
    X-bar theory, last I checked, is alive and well in HPSG. It has, however, been abandoned in recent versions of minimalism. – Alan H. Sep 13 '11 at 22:58
  • @AlanH.: As far as I understand HPSG, it lacks the X' node, X being a node rather than a leaf. – Evpok Sep 13 '11 at 23:12
  • 1
    HPSG and statistical models are not the same thing. "Surface" gets used in at least two ways: it can indicate a model which treats sentences as linear arrangements of strings (like many statistical NLP models), or it can be used to indicate a theory which is not derivational (i.e., has nothing like transformations). To my knowledge, most syntactic frameworks fall into the latter category, but I can't think of any which fall into the former. – Alan H. Sep 13 '11 at 23:15
  • 1
    Note from the future : not actually a X' parser but at least a Chomsky-Schützenberger one “Implementation of a Chomsky-Schützenberger n-Best Parser for Weighted Multiple Context-Free Grammars”/anthology/N19-1016.pdf – Evpok Oct 15 '19 at 9:29

The simplest answer is that high performance NLP applications do not use X-bar theory as an explicit representation. A major factor in this is that parsing is most commonly evaluated against the Penn Treebank or various dependency annotated corpora, neither of which would be considered X-bar structures. For an idea of what kind of structures and information are used in statistical parsing, check out one of the seminal articles such as Head-Driven Statistical Models for Natural Language Parsing (Collins, 2003).

There is certainly some work in NLP that does not explicitly use an X-bar representation but shows that constraints that could be considered to have come from X-bar (or equally, later from Minimalism) are useful in certain applications. For an example in parsing, see the Empty Categories section of Fully Parsing the Penn Treebank (Gabbard et al., 2006); for POS tagging see A Simple Unsupervised Learner for POS Disambiguation Rules Given Only a Minimal Lexicon (Zhao and Marcus, 2009).

Outside of performance-driven NLP applications, there is some work. As Alan H. points out, Sandiway Fong's Principles and Parameters Parser is a great example. It follows from Principle-Based Parsing (Berwick, 1987). For more background in the evolution of these systems, see Robert Berwick or Mitch Marcus's dissertations.

| improve this answer | |

I believe the problem with parsing X-bar structures is mainly performance based, it is difficult to get a parser to be fast enough to be useful. The transformations make the parsing non monotonic, which makes everything a royal pain for a computer.

There are however many systems who use other theories of syntax, such as HSPG, LFG, DCG, CG, TAG etc.

| improve this answer | |
  • Wouldn't the ambiguities also being a factor, leading to either multiplying numbers of parse trees, or potentially choosing the wrong parse tree? – hippietrail Sep 14 '11 at 10:37
  • Ambiguity is a problem for all NLP systems. Most parsers would deliver several (in the order of thousands) parse trees for sentences with more than about six words. However, stochastical models are used to try to guess the most probable one. I'm not sure if this task is more difficult for transformation based grammars, but I woudn't be surprised if it was. – user47 Sep 14 '11 at 10:40
  • Yes the more I use Google Translate the more apparent it becomes that stochastic models though very useful still allow all kinds of wrongness. – hippietrail Sep 14 '11 at 10:45
  • @johanbev: This is seriously misrepresenting the number of parse trees. There aren't a thousand parse tree for a seven word sentence. I would at best agree that there are "dozens" of parsings of many ADVERB/ADJECTIVE phrase sentences, but these are all different in where they bind the phrases. The real ambiguity is the same an in natural language reading. – Ron Maimon Mar 13 '12 at 4:10
  • X' structures are used in LFG, too. The performance is excellent with modern algorithms until constraints are evaluated. Ron Kaplan wrote a few papers on the combination of CF parsing with constraints. From the practical standpoint, however, X' rules are highly problematic and therefore avoided, as in the ParGram project (the problem is discussed at length in the book "Grammar writer's workbench"). – Atamiri Sep 22 '13 at 15:13

I believe that Sandiway Fong has made at least one principles & parameters parser, which might be old enough to have proper x-bar structure. There aren't many in that area doing computational linguistics, but probably the big name is Robert Berwick, so you might want to look at some of his work as well.

In response to another reply: Some popular versions of transformational grammars would be factorial time (based on number of words in a sentence), at least in the worst case, if they were implemented. Many who work in that line don't consider that important, because they don't consider what they do to be a model of performance but purely a model of knowledge. But not everyone buys that. See this article, for example. But I don't think the main culprit is x-bar theory. The main problem is global economy conditions, which in some models require the syntax to generate many permutations of a given sentence and then evaluate them after the fact.

| improve this answer | |

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.