Online Latent Structure Training For Language Acquisition

1

• Online Latent Structure Training For Language
  Acquisition
• Michael Connor, Cynthia Fisher and Dan Roth

Semantic Role Labeling
Experimental Results
Language Acquisition
Input
The dog chased the black cat
Task is to predict semantic roles given text as
input, treating identity of nouns and verbs as
hidden structure to be learned.

• Train a computer to learn language the way a
  child does
• Extract semantics and structure from text using
  ambiguous semantic feedback and simple
  representations based on defensible background
  knowledge.
• World does not provide complete indication of
  syntax or semantics, so how can children begin to
  learn?
• Must combine limited knowledge of both to be able
  to extract some meaning
• Develop online latent classifier that uses this
  to begin interpreting novel sentences

N V N
Hidden

• Trained on Child Directed Speech
• Test both ability to recognize argument and
  predicate and classify semantic roles
• Compare different levels of feedback
• Gold true arguments and roles
• Full full semantics, no arguments
• Set Unordered set semantics
• Super Superset of true roles

Feedback
A0 A1
(Agent)
Full, per word
(Patient)


Online Latent Structure Classifier
Role Classifier Features
Input Sentences with role feedback Output Two
linear separators for hidden structure,
for role classification For each
sentence

• Representation based on hidden structure to
  classify argument roles
• Linear order of Nouns and Verb
• Lexical Features
• Noun and verb word form
• noun'dog', verb'chase'
• Noun Pattern
• First of two, second of three, etc.
• dog is first of two
• Verb Position
• Dog is before the verb

Syntactic Bootstrapping
Recovering from Ambiguity

• Inspired by Structure-Mapping account of
  Syntactic Bootstrapping
• Children use very partial knowledge of syntax to
  guide sentence comprehension
• Does not assume verb learning must precede syntax
  learning
• Makes three key assumptions
• Sentence comprehension is grounded by the
  acquisition of an initial set of concrete nouns
• Children can identify referents of some nouns via
  cross-situational observation
• Children treat each noun as a candidate argument
• Number of nouns in the sentence is a cue to its
  semantic predicate-argument structure
• Children represent sentences in abstract format
  that permits generalization to new verbs
• Generalize early, once some nouns have been
  identified
• Incorporate this theory into model and
  demonstrate that a representation as simple as a
  set of nouns is useful as a foundation for new
  learning

Structure Classifier

• Superset feedback does not contain enough
  information to learn. But children probably
  couldn't guess at roles given scenes without
  knowing any nouns!
• Use bottom-up minimally supervised information to
  recognize arguments.
• HMM based noun and verb identification
• (Connor et al., ACL 2010)
• Uses only knowledge of small set of concrete
  nouns to seed both noun and verb identification
• Constrain structure and semantics in joint latent
  inference
• Only select possible interpretations where
  structure matches HMM prediction and corresponds
  to semantic feedback.

• Separate classifier for hidden structure given
  text
• Predicate Specific features
• Argument count
• Verb suffixes

Ambiguous Semantic Feedback
World does not provide true semantic feedback at
the level of roles per noun. Incorporate
alternative ambiguous semantic feedback through
constraints on interpretation in joint prediction.
Unordered Set
Superset
Input
Input
The dog chased the black cat
The dog chased the black cat
A0, A1
A0, A1, A2, AM-DIR
Feedback
Feedback
Conclusions
Implies number and label of roles
Contains true roles, plus random other
Possible Interpretations
Score
Score
Possible Interpretations

• Able to learn hidden structure from full semantic
  feedback
• With plausibly ambiguous semantics, cannot
  successfully learn hidden structure
• Learner needs to use both partial bottom-up and
  top-down information to successfully learn
• Number of arguments information is crucial,
  whether it comes from an informative scene, or
  from a partially parsed sentence

0.76
1.76
The dog chased the black cat
Input
The dog chased the black cat
Input
h y
V N N
Hidden
N V
Hidden
Prediction
Prediction
A1 A0
A1
1.63
0.50
The dog chased the black cat
The dog chased the black cat
Input
Input
h y
N V N
Hidden
N V N N
Hidden
Prediction
Prediction
A0 A1
AM-DIR A1 A2
The dog chased the black cat
The dog chased the black cat
Input
Input
1.13
0.93
N V N
Hidden
N V N
Hidden
Prediction
Prediction
A0 A1
A2 A0
Supported by NSF grant BCS-0620257 and NIH grant
R01-HD054448