Gambel

1
GambelYangs Mechanisms and Constraints in Word
Segmentation

• Virginia Savova
• Statistical Language Learning Reading Group,
• MIT 2005

2
Goals

• Evaluate statistical word segmentation strategies
• Establish the role of domain-specific (prosodic)
  constraints
• Propose an algebraic strategy based on a
  domain-specific but language-independent
  constraint

3
Critique of existing statistical word
segmentation algorithms

• Off-line, global optimization approaches
• overestimate learners computational capabilities
• Start from phonemes, rather than syllables
• underestimate what was learned on previous stages
• Discussion
• To what extent can off-line methods be converted
  to online
• Does assuming discrete learning stages make
  sense?

4
Corpus
CHILDES
PHONEME STRESS
SYLLABLES
DICTIONARY
MAX ONSET PHONOTACTICS
Test and train on the same corpus
5
Local Minima of Transitional Probabilities

• Formalization of Saffran et al.s hypothesized
  learning mechanism

TPab P(a b) P(a)

• Local minima Word Boundaries
• TPX-1gtTP X lt TPX1

6
Local Minima of Transitional Probabilities

• Results

• 23.3 recall
• 41.6 precision

• Problem monosyllables

• Discussion
• amend LM criteria?

Example monosyllable TP profile
7
Swingleys TP algorithm

• Probability of syllable, bigram, trigram
• Mutual information Transitional probability

Ixy log2 P(x y) P(x)P(y)
Ps/b/t F(s/b/t) n

• Fixed threshold ?
• If Ps/b/t gt ? then s/b/t is a word.
• If Ixy gt ? then xy is a word.

8
Swingleys TP algorithm

• Off-line
• Combines two measures in an arbitrary way
• Arbitrary cutoff at trigram
• Results
• high precision overall (?)
• low 3-syllable precision (23)
• low recall (27)
• ? TPs alone dont work

9
TP with Prosodic Constraints

• Unique Stress Constraint a word bears at most
  one (primary) stress
• universal (language-independent) constraint
• Strong (stressed) and weak (unstressed)
• (a) CONflict versus (to) conFLICT
• TPUSC
• S1S2 ? boundary
• S1W1...nS2 ? pairwise TP boundary (?)
• Results for TP USC
• 73.5 precision
• 71.2 recall

10
Algebraic learning with the Unique Stress
Constraint

• S1S2 ? boundary
• S1W1..i-1 ...Wj1..nS2? Wi...Wj is a word
• Add words to lexicon
• S1W1...nS2? split at random
• Results
• 95.9 precision
• 93.4 recall

11
Conclusions

• Local minima of TPs alone dont work
• Prosodic constraints help
• Universal prosodic constraint is sufficient
• Statistical learning is not necessary, algebraic
  does better