Language learning (acquisition): what is it?
- Presented with a sequence of grammatical sentences (or words),
be able to recognize novel sentences (or words) as
grammatical or ungrammatical.
- Presented with several instances of a word in context, be able to
interpret or produce that word appropriately in a new context.
- Presented with grammatical sentences in context,
be able to interpret or produce novel sentences in new contexts.
- Presented with sentences (words), learn to find the boundaries between the units
(words, morphemes).
Language learning: why is it hard?
- Language learners receive a finite amount of sometimes errorful
input (the "poverty of the stimulus" argument).
- Language learning is mostly unsupervised.
- Learners receive positive evidence, but no (direct) negative evidence.
- Learning is sometimes very rapid.
Kinds of information
- Two kinds of evidence in learning
Positive evidence | Negative evidence |
|
|
- Positive evidence
- Information about what is correct
- Causes the system to extend its hypothesis
- Negative evidence
- Information about what is not correct
- Explicit, labeled incorrect patterns
- Correction for the system's errors
- Information that there is an error somewhere, as in reinforcement learning
- Indirect (implicit) negative evidence: non-occurrence of something that would
otherwise be expected; useful in generative (Bayesian) learning models
- Active learning: learner asks about questionable
utterances, receiving useful negative feedback
- Causes the system to restrict its hypothesis
- Without negative evidence, is language learning possible?
- Alternatives to negative evidence: constraints
- Built-in (innate) constraints
- Constraints from the environment (input)
The problem for word meaning
- Is word learning supervised?
- Is negative evidence available?
- Quine's problem
- Figuring out what the referent of a novel word is
- Figuring out what aspects of the referent a novel word refers to
- Which word in an utterance goes with which part of a scene?
- How learning might be constrained
- Constraints on inter-lexicon relationships: mutual
exclusivity
- Constraints on kinds of categories (meanings)
- Whole object
- Taxonomic vs. thematic relations
- "Dimensional" (attentional) biases: shape, material;
language-specific biases (L. Smith, Bowerman, etc.)
-
Syntactic bootstrapping (Gleitman, etc.):
guess the meaning of new words on the basis of the meanings of words that
occur in similar syntactic patterns
Lois gorped the brick to Clark.
- Constraints from non-linguistic behavior
The problem for grammar
- Learning English passive
- The girl is hugging the woman.
The woman is being hugged (by the girl).
- Chuck chopped the onions.
The onions were chopped (by Chuck).
-
Lois became a lawyer.
*A lawyer was become (by Lois).
-
Semantic bootstrapping (Pinker): use (built-in) semantic features to constrain the kinds of syntactic patterns that are possible
- Grammatical parameters (Chomsky, etc.)
Innatist vs. empiricist theories
- What could be innate
- General purpose mechanisms (not specific to language)
- Language-specific mechanisms
- Categories (noun, agent, etc.)
- Parameters: set on the basis of input
- The importance of the input
- How impoverished is it?
- How does it change with the competence of the child?
- How variable is it (within and between languages)?
- How do children learning different languages differ?
- What statistical regularities does it embody?
- Empiricist possibilities
- The role of the environment: embodied models