Combining concepts

1
Combining concepts

• Cognitive Science week 9

2
compositionality

• Fuzzy set model
• Selective Modification model
• Semantic Interaction model
• CARIN model
• Dual-process model of noun-noun combination
• knowledge and pragmatic factors

3
This is too simple to work

• Dog tail barks wet_nose
• Red red
• red dog red tail barks wet_nose
• Why not?

4

• What does red modify the coat of the dog, its
  nose?
• What colour is red?
• red brick, red wine, red pillar box
• Compounds
• red lurcher
• sandy fawn red lurcher http//www.doglost.co.u
  k/forum.asp?ID9757

5

• Red is an intersective adjective
• Extensionally, simple set intersection almost
  works (apart from the problems above)
• Skilful set intersection simply wont work
• Betty is a skilful ballerina, but shes useless
  at rugby.

6
Fuzzy set theory

• Instead of True (1) or False (0)
• shades of gradable truth 0, 1
• Eg. A showjumper is a jockey 0.7
• Use a rule to combine these

7
Red jockey

• Take some object
• Lets rate it as a jockey 0.7
• as a red thing 0.8
• The rule is min, take the minimum
• As a red jockey, it should be 0.7

8
Conjunction effect

• He would typically be rated as a better instance
  of red jockey
• than of red or jockey
• Another example, a brown apple
• This is contrary to the min rule

9
Selective Modification model

• Represent concepts as frames
• a set of slots with potential values
• each slot is weighted (salience)
• Apple 1.0 COLOR red 25
• green 5
• brown
• 0.5 SHAPE round 15
• square
• 0.3 TEXTURE smooth 25
• bumpy

10
Selective Modification model

• Goodness measured by adding up matches (and
  taking away mismatches)
• Object (X, COLOR brown, SHAPE round, TEXTURE
  smooth)
• Apple 1.0 COLOR red 25
• green 5
• brown
• 0.5 SHAPE round 15
• square
• 0.3 TEXTURE smooth 25
• bumpy

1.0 0
0.5 15
0.3 25 15
11
Selective Modification model

• Combination selects slots
• disambiguates potential values
• increases weight of selected slot
• Apple 1.0 COLOR red 25
• green 5
• brown
• 0.5 SHAPE round 15
• square
• 0.3 TEXTURE smooth 25
• bumpy

Red
12
Selective Modification model

• Combination selects slots
• disambiguates potential values
• increases weight of selected slot
• Apple 2.0 COLOR red 30
• green
• brown
• 0.5 SHAPE round 15
• square
• 0.3 TEXTURE smooth 25
• bumpy

Red
13
Selective Modification model

• Combination selects slots
• disambiguates potential values
• increases weight of selected slot
• Apple 1.0 COLOR red 25
• green 5
• brown
• 0.5 SHAPE round 15
• square
• 0.3 TEXTURE smooth 25
• bumpy

Brown
14
Object (X, COLOR brown, SHAPE round, TEXTURE
smooth)

• Combination selects slots
• disambiguates potential values
• increases weight of selected slot
• Apple 2.0 COLOR red
• green
• brown 30
• 0.5 SHAPE round 15
• square
• 0.3 TEXTURE smooth 25
• bumpy

Brown
1.0 30
0.5 15
0.3 25 45
15
Selective modification too narrow

• Medin Shoben
• wooden spoon v. metal spoon
• brass, silver, gold coins? railings?
• Which pair is more similar?

16
Limits of Medin Shoben

• 1. What about lexicalisation?
• wooden spoon familiar, stored
• 2. What about ambiguity?
• gold1 made of the substance gold
• gold2 painted a gold colour
• 3. Lack of an explicit model

17
Semantic Interaction Model

• Dunbar, Kempen Maessen (1993)
• Property ratings
• nouns some peas
• adjective-noun some mouldy peas
• Effect of the adjective the difference
• Effect not the same for different nouns

18
Semantic Interaction Model
Some mouldy peas
Adjective-noun rating (target)
Noun rating (training input)
Some peas
19
Semantic Interaction model

• Results for adjective mouldy
• Training items broccoli .013
• cabbage .007
• bananas .001
• peas .027

Test item carrots .011 Mean error for carrots
with random weights (10 runs) 0.49
20
Noun-noun combination

• peanut butter butter made of peanuts
• mountain hut hut in the mountains
• zebra bag bag with zebra pattern
• Property v. relational interpretations

21
CARIN model

• Gagne Shoben (1997)
• Past patterns affect interpretation
• (cf. statistical models of disambiguation)
• People interpret faster if the relation is one
  that has often been used with this modifier
• Eg. football scarf, football hat ? football flag

22
CARIN model

• Created a corpus of novel NN combinations
• Judged interpretation for each NN
• Counted frequency of different kinds of
  interpretation for each N
• Used frequency to predict
• Timed judgement does this NN make sense

23
Dual process model (Wisniewski, 1997)

• relational
• the modifier occupies a slot in a scenario drawn
  from the conceptual representation of the head
• property (and hybrid)
• Two-stage process
• 1. Compare areas of similarity, so difference.
• Differences - candidate for the property to move
• Similarities - aspect to land the property on
• 2. The property transferred is elaborated.
• NN combinations are largely self-contained, a
  function largely of "knowledge in the constituent
  concepts themselves" (1997, p. 174)
• discourse context may influence

24

• Wisniewski's evidence includes participant
  definitions for novel combinations presented in
  isolation
• property mapping as well as thematic
  interpretations (Wisniewski, 1996, Experiment 1)
• property mapping is more likely if Ns are similar
  (Wisniewski , 1996, Experiment 2)
• novel combinations
• null contexts
• "listeners have little trouble comprehending
  them" (Wisniewski, 1998, p. 177)

25

• In real-world lexical innovation there is an
  intended meaning
• Conjecture
• The need to convey an intended meaning, rather
  than only the ability to construct a plausible
  interpretation, is key to understanding NN
  combination in English. NN combination is
  primarily something the speaker does with the
  hearer in mind, rather than the converse.

26
Pragmatics - Relevance

• Sperber Wilson (1986)
• Principle of Relevance presumption that acts of
  ostensive communication are optimally relevant.
• Optimal relevance
• 1. The level of contextual effect achievable by a
  stimulus is never less than enough to make the
  stimulus worthwhile for the hearer to process.
• 2. The level of effort required is never more
  than needed to achieve these effects.

27
Pragmatics - Relevance

• Speaker chooses expression that requires least
  processing effort to convey intended meaning.
• Consequently, first interpretation recovered
  (consistent with the belief that the speaker
  intended it) will be the intended interpretation.
• If first interpretation not the correct one, then
  speaker should have chosen a different
  expression, for example by adding explicit
  information.

28
Clark and Clark (1979)Denominal verbs -
"contextuals"

• Tom can houdini his way out of almost any scrape
• Sense can vary infinitely according to the mutual
  knowledge of the speaker and hearer
• Any mutually known property of Houdini, if
  speaker
• "... has good reason to believe... that on this
  occasion the listener can readily compute the
  intended meaning ... uniquely... on the basis of
  their mutual knowledge..."

29

• Pragmatic approaches emphasise cooperative and
  coordinated activity by both speaker and hearer.
• Self-containment approach emphasises NN
  combination as a problem for the listener.
• On pragmatic account, notion of an interpretation
  in isolation from any context is defective

30

• Prediction
• ? readers presented with novel stimuli in
  isolation will experience difficulty
• They cannot make the presumption of optimal
  relevance, since they have no evidence of
  intentionality
• They therefore have no basis for differentiating
  the intended interpretation from any conceivable
  interpretation.

31

• A simple experiment can participants interpret a
  novel NN in isolation?
• Key finding
• Participants were typically unable to provide the
  correct interpretation.
• In addition, they knew they didnt know.
• See Dunbar (2006) for details.

32
Review

• Fuzzy set model
• Selective Modification model
• Semantic Interaction model
• CARIN model
• Dual-process model of noun-noun combination
• knowledge and pragmatic factors