1 the logic of default inheritance in word grammar richard hudson lexington, may 2012
Post on 19-Jan-2018
217 Views
Preview:
DESCRIPTION
TRANSCRIPT
1
The logic of default inheritance in Word Grammar
Richard HudsonLexington, May 2012
2
Plan
1. Word Grammar 2. WG morphology3. Five problems for DI, and the WG
solutions4. Conclusions
3
1: Word Grammar
• Aims– to model language in cognition
• 'graceful integration'
– to discover if anything is unique to language• Assumptions
– cognition is a network– the network includes a taxonomy– the model must include processes as well as structures.
4
WG and language structure
• Language is a network• The network links sounds to meanings
– but NOT directly– so the units are not 'symbols' or 'constructions'
• The network defines:– 'levels'– various sub-taxonomies
5
'Levels' in a network
realization
pronunciation
morph
spelling
sound letter
concept
meaning
semantics
syntax
morphology
phon/graph-ology
word
morph
sound
6
Taxonomy in a network
concept
entity relation
letter wordmorphsound realization
pronunciation spelling
meaning
'is-a'
7
WG and processes
• Spreading activation– including attention
• Node-creation for tokens/exemplars– e.g. word-tokens in recognition or planning
• Binding– e.g. token to token in parsing
• Default inheritance
8
An example outside language: hearing some purring
cat
token 1
token 2?
binding
node-creation
purring
activationinheritance
token 3
purrer
9
An example inside language: hearing /kat/
CAT
realization
token 1
token 2?
/kat/
token 3
10
2: WG morphology
• Inferential and realizational• Morphosyntax
– words realized by meaningless morphs– any degree of generality
• word-class – lexeme – sub-lexeme
• Morphophonology– morphs realized by phones or letters
11
Morphosyntax in a network
word
adj-of
noun
BOYISH BOY plural
{boy}base
full
s-variant
inflectional morphology
A
full
base
adjective
derivational morphology
12
Morphophonology in a networkword {z}
bases-variant
{z}
1
2
pronunciation
/z//ɪ/
210 #
/s//ɪ/+1 #
sibilant
next
voiceless
next
13
3: Five problems for default inheritance
• The basic idea is very simple:– defaults generalise except when overridden
• But there are challenges:a. generality: going beyond morphologyb. reliability: non-monotonicityc. certainty: recognising clashes and winnersd. relevance: avoiding irrelevant enrichmente. economy: avoiding storage of outcomesf. sensibleness: avoiding silly categorization
14
DI in WG
• Since 1980 I've tried several different theories of DI– e.g. 1990: exceptions are marked with 'NOT …'
• 2007: A solution:– Default inheritance only applies to tokens. – DI is driven by spreading activation.– The network defines conflicts and outcomes.
15
a: Generality
• Default inheritance is part of general cognition.– Hence: prototype effects
• So the mechanism used for morphology must generalize – beyond ready-made AVMs– e.g. 'birds fly' but 'penguins don't fly'.
16
Generality in WG
• Properties are links to other concepts– not AVMs– e.g. 'Birds fly' is a property of 'bird' as well as
of 'fly'. – 'Not P' conflicts with 'P' and overrides it.
• 'not P' in a network: 'quantity of P = 0'
17
Penguins
bird
flier
flying
penguin
token
0#
0#
18
b: Reliability
• Monotonicity: inferences are– cumulative– reliable – i.e. not liable to later retraction
• But DI is non-monotonic– because default properties may be overridden– unless special ordering restrictions are imposed– but this is odd in a declarative system.
19
People or persons?noun
plural
full
s-variant
PERSON
PERSON, plural {people}
full
allows persons
20
Reliability in WG
• Inheritance only applies to tokens– i.e. at the foot of the taxonomy
• And it applies recursively working upwards– so the first value inherited is always the winner.
• So DI is, in fact, monotonic!• And DI is always part of the process of
node-building.
21
People, not personsnoun
plural
full
s-variant
PERSON
PERSON, plural {people}
full
token
22
c: Certainty
• Inheritable properties may be uncertain:– Multiple inheritance allows two outcomes.– A node's identity may be defined
• directly by a 'filler'• indirectly by reentrancy.
• In conflicts, which property wins?• WG: let the network decide.
23
Multiple inheritance
• No special provisions needed.• If conflicts arise, so be it.
– e.g. *I amn't• = BE, 1sg, negative• 1 sg: realization = {am}• negative: realization = {aren't}• unresolved because 1 sg and negative are sisters.
• Reentrancy (loops) is harder.
24
Is people the s-variant of PERSON?noun
plural x
full
s-variant
PERSON
PERSON, plural {person}
{z}2
1token y
Does token inherit s-variant?
If {person} isa x,
… and the consequences!
But if not,
then y inherits s-variant.
all is well.
25
Certainty in WG
• If: the inheritor for token x finds [X, R, Y]• = [X R Y] or [Y R X]• 'upper-case' dominates 'lower-case'
• And if:– x already has [x, r, …]
• then: Do nothing.– x already has [x, …, y]
• then: create [X, r, y], [r isa R]– otherwise: create [x, r, y], [r isa R], [y isa Y]
26
Certainty in a network
x
R
X Y
If: [X, R, Y]
and if: already [x, r, …]
Otherwise:create [x, r, y], [y isa Y], [r isa R]
Then: Do nothing.
Then: create [X, r, y], [r isa R]
Otherwise if: already [x, …, y]
and X dominates x
y
r
27
d: RelevanceThe inheritance explosion
• Each new node created is a token– so it inherits too, creating further nodes
• Every node in the taxonomy is an inheritance source.
• Every property can be inherited.• Inheritance is resource-intensive.
– E.g. 'Does a canary have skin?' takes longer to answer than 'Does a canary sing?'
• So we need a rationing mechanism.
28
Relevance in WG
• Do we inherit irrelevant properties?– e.g. spelling– etymology
• Relevance in WG:– only inherit active properties.– Activation levels vary, and reflect:
• past experience• present concerns
29
Context-dependent DI
school SCHOOL
etymology
skhole
meaning
token
Assume: Total inheritance.Relevant inheritance.
30
e: Economy
• Do we store inherited properties?– Not in general.– Mixture of full and empty entries.
• Economy in WG: Only tokens inherit– So types aren't enriched.– But tokens may become types
• i.e. may be 'learned'
31
Fringe activityMemorychanges slowly
Experiencechanges
fast
32
f: sensibleness
• Problem: How to avoid silly classification?– e.g. a block of wood isa bird, but overrides all
bird properties.• Solution: A theory of learning and use:
– We only build isa links where properties are shared:
• in creating new category nodes• in creating new token nodes.
33
5: Conclusions
• DI is part of cognition.– and needs a cognitive model– it mustn't be limited to AVMs.
• It's part of token-building.– so it's monotonic.
• It's limited to active relations.– so it only inherits what's relevant.
34
Thank you
• This slideshow can be downloaded from:www.phon.ucl.ac.uk/home/dick/talks.htm
• For more on Word Grammar, see:www.phon.ucl.ac.uk/home/dick/wg.htm
top related