Sunday, November 15, 2009

Construction grammar

The subj (CxG) is exactly what I had in mind. In short:

  • There's no language-specific unit in the brain, the language ability is due to common cognitive units.
  • Basic notion in the language is a construction (a template text fragment). No one has defined it but everyone uses.
  • Both the speaker and the hearer operate in terms of constructions, not words. Unlike other prominent theories.
  • Children learn specific constructions first, abstracting them into more general rules afterwards.
  • Lexical items which are often used in similar contexts tend to grammaticalize into new constructions as language evolves.

    Construction approach seems very promising to me, though I see at least two weaknesses in it, to which I haven't found an adequate answer so far:

  • Why learning new languages becomes much harder after puberty. Is it a degradation in common cognitive abilities? What's different with second language learners? Why do Japanese speakers miss 3d singular -s in English ('he like driving')?

  • CxG accounts only for positive data, and doesn't explain why exactly the ungrammatical samples are so (or not so) ungrammatical. A vague explanation that one gets used to a specific way of expressing an idea and all the other ways are not so habitual hence more difficult to produce/analyze, doesn't satisfy me very much. It may be equally difficult (or easy) from processing perspective. E.g. adjectives in English would have come after the noun with no clear performance disadvantage. Semantics could also be clear, like in 'he go cinema'.

    Another explanation could be an Occam's Razor. In all ungrammatical examples I've seen, there is a grammatical counterpart. So, the brain could think, why should the same meaning be produced in this strange way while there's another well-established one, and mark the 'strange way' as an ungrammatical.

  • The question remains, how to create a parser based on constructions, and how to induce those constructions from a corpus. And, as usual, what that parser should produce as the result.


    Дана said...

    I like this post.
    And you're right about some weaknesses of the CG.

    Anyway, there is no grammar yet that could explain the difficuties the adult L2 learner faces. :)

    By the way, CG is good at explaining why why L2 learners are better at language acquisition at the early steps - they don't need to form some basic grammatical categories such as a noun, a verb, etc. If I remember correctly, chomskian syntaxists (at least Pinker)assume that basic cognitive categories corresponding to the noun and verb already exist in the child's mind before he starts acquiring a language.

    As for grammaticality problem, the CG rules it out without any problem, as for me.

    Take a subject-verb agreement in English:

    REF1 + VERB + REF2

    If REF1 is "I", "you", "we" "they" or "Noun + Noun" (= "Mary with Sarah", "Mummy and Daddy") > then do not add anything to V in Present tense.

    If REF1 is "Noun" or "he/she/it" - then add -s.

    There is little chance that a child can be exposed to many sentences violating the rules (except for the case when he/she lives in a immigrant family and doesn't see anyone except for his parents).

    As for grammar features that are less salient in the input - that's really a problem for a CG.

    Peter Gromov said...

    Some linguists (e.g. Croft) claim that there are no such cross-lingustic word classes as noun or verb at all. And Evans&Levinson say that Straits Salish language has only verbs, no nouns. CxG handles this very well saying that there just constructions which need not correlate to any innate word classes.

    L2 learners may nevertheless apply abstract constructions from L1 and succeed at some time if L1 and L2 are similar in some aspect.

    Subject-Verb agreement violations do indeed seem not so ungrammatical for me, they just give a flavour of 'they don't say it this way and therefore it's incorrect'. There are more severe cases like wh-movement constraints violations when it's just hard to understand what this could mean ("*Who did Brian know why Tom met?"). Anyway long-distance dependencies seem to be a major problem for CxG.

    I just wonder, what the grammar features that are not salient in the input can be? If they are not in the input, they're not in the output. That's what I've thought.

    Дана said...

    Ok. I agree that advantage with classes as a whole is not so good to be applied here. I just would like to add that sometimes L2 learners know something that they aren't supposed to know from input from their native language (e.g. Chinese people don't make errors in pronoun case in speeded grammaticality judgement task, although they commit errors with subject-verb agreement; the range of examples is quite wide here). Thus, their grammatical knowledge can't be usage-based at all its aspects. There must be some deeper layers of language competence that are in charge for this.

    As for the degree of ungrammaticality of sentences with subject-agreement violations, do they still seem to you not so ungrammatical when they occur in pro-drop languages? :) Anyway won't hearing a sequence of sentences with subject-agreement violation in Russian make your brain explode soon or suffer from the amount of ELAN it gets? :)

    I don't understand why long-distance dependencies are a problem for CG. They require exactly the same procedure of segmentation, only the units are larger. Or you mean, the CG can't explain why the wh-island constraints are rarely violated by L2 learners?

    As for your last point, a good example of rules that can't be fully predicted from the input is Russian aspect. Children don't make errors with it, although they can hear both verbs of perfective and imperfective aspect in similar contexts like:

    - Я всю неделю ждала/прождала его звонка.
    - Я несколько раз запретил/запрещал ему делать это.

    -В течение недели я ждала/*прождала его звонка
    -Я многократно запрещал/*запретил ему делать это.

    Natalia Gagarina says that children initially use aspects in prototypic cases and prefer perfective aspect in past tense contexts to avoid this ambiguity. So they don't master the rule in the full scope. But no errors - it sounds impressive, right? And the reason for which they don't make errors in such cases remains unexplained by the input they get. Remember, that marking of perfective aspect varies very much across Russian verbs, so the children can't rely on morphological form.