A network model of processing in morphology

1
A network model of processing in morphology

• Dick Hudson
• UCL
• www.phon.ucl.ac.uk/home/dick/home.htm

2
Plan

• The theoretical framework Word Grammar
• A tiny challenge find a word
• A small challenge take account of context
• A theory of processing
• A modest challenge do morphology
• A fair challenge plan a word
• A research strategy

3
Word Grammar

• The whole of language is a network.
• Links as well as nodes are classified by is-a
  links.
• Is-a links allow default inheritance.
• For example, the word CAT

4
(No Transcript)
5
A tiny challenge

• How do we work out that a word pronounced /kat/
  means Cat?
• I.e. how do we use a network to connect /kat/ to
  Cat?
• An uncontroversial answer by spreading
  activation
• Spreading activation explains
• Priming (e.g. doctor primes nurse)
• Speech errors (e.g. doctor replaces nurse)

6
From /kat/ to Cat

• Activation spreads blindly
• from /kat/
• via CAT
• to Cat
• Like this

7
(No Transcript)
8
and back?
9
The research questions

• Q1. Are comprehension and production different
  processes?
• Q2. Is lexical processing different from
  morphology (and syntax and )?
• Q3. Exactly how does spreading activation help?
• Q4. How does context contribute?

10
Q4. Context a small challenge

• How do we work out that /bat/ means winged
  mouse when were thinking of winged mice?
• The winged-mouse concept is already active, so it
  becomes more active than cricket-bat.

11
What kind of bat?
12
So what?

• A model of processing must be
• Interactive, allowing information to interact
  freely
• NOT modular, limiting flow of information
• A model of knowledge must be
• Integrated, with many links between language and
• general knowledge
• contextual knowledge

13
Q3. How spreading activation helps

• Spreading activation is necessary.
• But not sufficient. Processing also requires
• Creation of new nodes.
• Guidance provided by the new nodes.

14
New nodes for tokens

• Every token needs a new node because
• Its a distinct concept which we monitor and may
  remember.
• It may even be deviant, e.g. mispronounced.
• So we create a new node for each token.
• E.g. we hear /kat/ and create X.

15
What we know about tokens

• A token is partially known.
• E.g. For X we know
• X has /k/ /a/ /t/ as its parts.
• X is-a Y (not known).
• Y is-a form.
• The known properties are the source of the
  processing.
• Theyre highly active because recent.

16
What we dont know about tokens

• We dont know some properties.
• Hearing we dont know unobservable properties
  (meaning, etc).
• Speaking we dont know observables.
• We want to know some of these properties.
• So these nodes are also highly active.
• These properties are the target.

17
Activation is guided

• Therefore activation always spreads from at least
  two nodes.
• The source
• already enriched and active.
• The target
• Needs enrichment.
• These two nodes guide activation intrinsically.
• So we dont need extrinsic control.

18
How to get rich.

• Impoverished nodes need to become rich.
• How?
• By marrying a rich node.
• This is binding poor node rich node
• An active token variable binds to its most active
  sister.
• For example, Y binds to cat.

19
From /kat/ to cat
20
The best fit principle

• An impoverished node binds to its most active
  sister.
• NB spreading activation is global,
• activation may come from any part of the network
• This guarantees the best global fit.
• So context can override form
• We recognise ltthatgt as a wrong ltthangt.

21
Inheritance

• Active token nodes
• Bind to the most active sister(s).
• Inherit from their mother(s)
• So

22
After the wedding .
23
The inheritance!
24
An algorithm for processing

• Add active nodes for tokens.
• Spread activation.
• Bind active token variables to constants.
• Inherit to active tokens.
• And do it all again.

25
So what?

• Spreading activation has to interact with
  node-creation
• not found in most other models.
• Processing in a network changes the network nodes
• Not just their activity levels

26
Selective binding

• When we hear a word, we look for
• its sense
• NOT its etymology or its French translation
• Unless were discussing etymology.
• Why?
• Because the sense link is usually more active.
• Why?

27
Why not etymology?
28
Links between links

• Every link is-a some more general link.
• So activation can flow from link to link.
• So some links start more active than others.
• So we concentrate processing resources on
  interesting properties.

29
Q2. How about morphology (etc)?

• How do we recognise /kats/?
• Just the same procedure.
• But the processor activates two models
• CAT, via /kat/ and cat
• Plural, via /s/ and s
• This requires multiple default inheritance.

30
(No Transcript)
31
So what?

• Complex combinatory patterns can be
• inherited.
• bound.
• If it works for morphology, maybe it will work
  for syntax too?

32
Q1. Is production the same?

• Yes.
• But the target and source are reversed.
• E.g.
• Source Cat
• Target ? ( /kat/)
• The direction of processing is decided by the
  choice of target, not by the structure of the
  system.

33
A research strategy

• How to go beyond hopes, promises and faith?
• Build a computer model of a processor.
• Add a database for a linguistic network.
• See if it works.

34
Meet WGNet
35
The team

• The programmer Geoff Williams
• The consultant Sean Wallis
• The linguist Jasper Holmes
• The sponsor ESRC

36
(No Transcript)
37
(No Transcript)
38
Morphology
39
(No Transcript)
40
Inheritance
41
Conclusion

• Maybe all mental interaction with the world
  follows the same principles
• Spreading activation.
• Default inheritance.
• Binding according to the Best Fit principle.

42
Thank you

• www.phon.ucl.ac.uk/home/dick/home.htm