Why it is Hard to Label Our Concepts

1
Why it is Hard to Label Our Concepts

• Jesse Snedeker and Lila Gleitman
• Harvard and U. Penn

2
Phenomenon to Explain

• 1st 50 words of acquired language massively
  over-represent nouns (45) compared to verbs
  (3).
• Those verbs that do appear are linked to action
  terms (e.g. go, run, throw) even though verbs
  like think, know are frequent in maternal speech
• Why is this?
• problem of conceptual development or linguistic
  learning

3
SG Proposal

• Language acquisition is mapping procedure
• Extralinguistic info (e.g. perceptual input) only
  suitable for mapping to concrete concepts
• Word-to-World pairing
• Verbs require bootstrapping from previously
  acquired concrete nominals
• Structure-to-World pairing (or syntactic
  bootstrapping)
• Main idea learners use multiple (different)
  converging cues to acquire vocab

4
The Human Simulation Paradigm

• 24 most frequent nouns/verbs in transcript of
  infant directed speech videos
• Video played to adult subjects without audio
• Tone played at moment target word is uttered
• Clip played 30 secs before tone 10 sec after
• 6 clips for each target word (diff. contexts)
• Told whether noun/or verb
• Subjects asked what the target word was

5
Experiment 1 nouns vs. verbs

• Identified 45 of freq nouns 15 of verbs
• Why?
• Early nouns are more concrete than verbs
• Learning highly correlated with imaginability
• Conclusion
• because limited to only extralinguistic data,
  could only get concrete words

6
Exp 2 affects of diff info types

• Paradigm change (slightly)
• Only verbs
• Multiple Conditions
• Scenes same as exp 1
• Nouns given lists of nouns/pronouns in target
  verb construction (no visual info)
• Frames given target verb frame with nonsense
  nouns (no visual info)
• SN see above
• SF see above
• Full see above

7
(No Transcript)
8
(No Transcript)
9
(No Transcript)
10
(No Transcript)
11
Interesting Ideas

• Language acquisition as mapping problem separate
  from conceptual development
• Split up context into features
• Suitable for ML techniques
• Pred/Arg struct matches Grammar
• is innate