1 / 41

How to learn and use a language

How to learn and use a language. Dick Hudson Cambridge, June 2008. Linguistic theory in 1960. Limited scope phonology, morphology diachronic and synchronic little syntax, little semantics no sociolinguistics, no psycholinguistics Competing theories about structure

ivy
Download Presentation

How to learn and use a language

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. How to learn and use a language Dick Hudson Cambridge, June 2008

  2. Linguistic theory in 1960 • Limited scope • phonology, morphology • diachronic and synchronic • little syntax, little semantics • no sociolinguistics, no psycholinguistics • Competing theories about structure • But little theory about fundamentals • e.g. ‘Where is language’?

  3. Linguistic theory in 2008 • Even more competing theories of structure • But these theories have much broader scope • include syntax and semantics • and fundamentals, e.g. I-language is cognition • Evidence from cognitive science is relevant: • neuroscience • cognitive psychology • artificial intelligence

  4. Neuroscience in 2008 • The brain is active. • Different areas are active at different times. • These differences relate to the architecture of ‘the mind’, including language. • So they must have implications for linguistic theory. • E.g. Is language a distinct mental organ?

  5. Where is language? (Roelofs 2008)

  6. Cognitive psychology in 2008 • The mind is a network • Evidence: spreading activation • Activation spreads blindly • So it activates all neighbours equally • So non-target nodes become active • and remain active for a while

  7. A mental network Reisberg 2007

  8. Network structure in language • Language must be a network too because it carries spreading activation. • Evidence for spreading activation: • Priming: word 1 primes word 2 if they are network neighbours. • Speech errors: a neighbouring word is accidentally activated.

  9. A priming experiment

  10. On the screen: no priming lorry fon nurse Delay: 0.9 0.8 0.8 Non-word Word

  11. Nurse primes doctor. doctor fon nurse Delay: 0.9 0.8 0.6 Non-word Word

  12. Priming at all levels Words prime network neighbours in: • Phonology: verse primes nurse (but only briefly) • Morphology: hedges primes hedge for longer than pledge does. • Syntax: Vlad brought a book to Boris primes other V + DO + PP sentences • Semantics: nurse primes doctor.

  13. Errors at all levels Errors involve either neighbours in: • Phonology: orgasms (organisms) • Morphology: slicely thinned (thinly sliced) • Syntax: I’m making the kettle on • For: making some tea + putting the kettle on • Meaning: crosswords (jigsaws) or an active mental environment: (Addressee is sitting at a computer.) You haven’t got a computer (screwdriver) have you?

  14. AI in 2008 • Generalisations apply by default inheritance • If A is a (‘isa’) B, • and B has property P • then A inherits P • unless P is blocked by a known property of A. • Allows exceptions • messy logic of everyday thinking

  15. An inheritance hierarchy Luger & Stubblefield 1993

  16. So what (1)? • Language is a network • and nothing else? • When speaking and listening we use: • spreading activation for retrieving • default inheritance for generalising • A linguistic theory should say so too • e.g. Word Grammar

  17. Word Grammar: networks Language is a network of atomic nodes • not a network of ‘lexical items’ or ‘constructions’ • the labels are irrelevant • all that counts is the links among nodes For example …

  18. purring KITTEN PURR CATTLE CAT CATMINT {cattle} {cat} /kat/

  19. Word Grammar: isa hierarchies The network includes a number of ‘isa’ hierarchies • These allow inheritance logic For example …

  20. mammal purring noun KITTEN PURR CATTLE CAT CATMINT {cattle} {cat} /kat/

  21. Word Grammar: structured links The links are structured by: • a direction • a classification • (actually, using ‘isa’) For example, ….

  22. mammal offspring purrer purring noun meaning KITTEN PURR CATTLE CAT CATMINT realisation {cattle} {cat} /kat/

  23. But … • Spreading activation • needs a directed database • so production has a different database from perception • Default inheritance • clutters the database • is non-monotonic (allows false inferences) !

  24. The solution: linguistic theory • Distinguish tokens from types • cf performance and competence • Why? • tokens have unique properties (time, etc) • tokens may be deviant (e.g. mispellings) • tokens are hyponyms of types

  25. Phrase structure black cat meaning ??? black cat ‘black cat’ meaning CAT ‘cat’

  26. Dependency structure black cat meaning ??? ‘black cat’ meaning CAT ‘cat’ adjunct meaning ‘black cat’

  27. Tokens bind • We use token nodes to ‘bind’ different properties together. • Colour, shape and size • Pronunciation, grammar, meaning, speaker, etc. “Nodes that fire together, wire together.” (Hebb 1949)

  28. Enrichment by inheritance • We recognise properties • observed (perception) or known (production) • We create a new node to bind them • We make it very active • Activation spreads from observables to the permanent network • This selects the best-fitting available type • Then we inherit from this.

  29. Directing activation • The active token node combines • a source: the known properties • a target: the unknown properties • Inheritance finds the unknown • So activation spreads automatically from known to unknown.

  30. Hearingcat • Hear /kat/ • Create an active node for a word token pronounced /kat/ • Spread its activity • Find the most active word type • Inherit a meaning, syntax, etc. • Integrate the meaning

  31. Sayingcat • Think ‘cat’ • Create an active node for a word token meaning ‘cat’ • Spread its activity • Find the most active word type • Inherit a pronunciation, syntax, etc. • Say /kat/.

  32. From /kat/ to ‘cat’ by inheritance inherited meaning meaning listening CAT word token pronunciation pronunciation /kat/ known

  33. From /kat/ to ‘cat’ by inheritance meaning meaning word token CAT speaking pronunciation pronunciation /kat/

  34. Who inherits? • Only tokens inherit. • but recursively up the isa hierarchy • Consequences for the logic: • it’s clean (reliable, monotonic) • it doesn’t clutter the database

  35. Only active relations inherit • Relations have activation levels too. • determined by attention and interest • We inherit only for active relations. • e.g. we can listen to words for: • meaning (normal) • words containing /b/ (experimental subject) • syntactic peculiarities (syntactician)

  36. A word has many properties meaning dependent etymology word rarity spelling pronunciation

  37. Learning • Remembering: some salient tokens become permanent types • NB tokens and types have exactly the same formal properties • All of language can be learned from experience of ‘usage’. • Generalising: by induction from specifics

  38. Induction by node-creation • We generalise when properties correlate. • if X and Y and … all have P1 and P2 • then an ‘X or Y or ….’ has P1 and P2 • Correlations emerge from activation: • P1 and P2 receive activation together from multiple sources. • “Nodes that fire together wire together…” • …by creating a new node for a super-category • Maybe this happens in down-time, not when processing?

  39. Inducing birds robin movement flying feathers cover bird feathers cover sparrow cover movement movement flying

  40. So what (2)? • Language requires: • a network of atomic nodes and links • spreading activation • multiple default inheritance • node creation • The same is true for other knowledge • So language is: • ‘just knowledge’ • a very clear window into the mind.

  41. Thank you • This slide show is available at: • www.phon.ucl.ac.uk/home/dick/talks.htm#cam • More about Word Grammar: • www.phon.ucl.ac.uk/home/dick/wg.htm • Language Networks. The New Word Grammar (OUP 2007)

More Related