the evolution of speech segmentation: a computer simulation
DESCRIPTION
These are the slides for my undergraduate dissertation on word segmneTRANSCRIPT
The Evolution of Speech Segmentation
A Computational Simulation
Richard Littauer (Edinburgh)
Outline
• The Problem
• The Possible Solution
• Conclusions and Implications
The Research Problem
• Word Segmentation
The Problem
• Fluent listeners hear speech as a sequence of discrete words.
• But there are no pauses in the wave form…
The Problem
• Listeners Problem:
• jakɑrəmnə (or thereishope)
• Solution!• Find all boundaries• Don’t find any boundaries
The Problem
• Suggestions:– Allophonic variation– Coarticulation– Prosody– Phonotactics– Combining any of these– Or…
The Problem
• Recent studies have shown that 8-month-olds can segment continuous strings of speech syllables into word-like units using only statistical computation of syllables (Aslin et al. 1997, 1998; Mattys et. al, 1999)
The Problem
• These studies looked at syllable transition probability, but didn’t look at the possibility that the children may simply be counting the syllables.
The Problem
• Furthermore, while Aslin, Saffran, & Newport (1996; 1998) did show that children can use statistical probability, they didn’t judge how that type of analysis would influence language over time.
The Problem
• No one has done this (as far as I am aware.)
The Problem
• So, why does this matter? Because, obviously, the child has no lexicon to back up, so the information which the child is exposed to must be that which is used to learn how to segment properly.
My Simulation
• Code for four different possible transitional segmentation strategies. Use an Iterated Learning Model to see how well these do when culturally replicated.
My Simulation
• Coded four different types of methods:– If you have seen one of the two test words before
and not the other, choose the one you have seen before.
– If one of the test words has occurred more frequently than the other, chose the more frequent one.
– If one of the test words contains more frequent transitions, chose that one.
– If one of the test words contains more probable transitions, chose that one
My Simulation
• Variables:– word recognition– word frequency– syllable transition count– syllable transition probability
My Simulation
• Variables:– Word length– Amount of ‘syllables’– Amount of words– Amount of words used– Fixed lexicons
My Simulation
• Types of Pairings– 2 randoms– 1 lex, 1 random word with the same phonemes– 1 scrambled word– 1 chopped up
My Simulation
• The ILM– All of this was run through an Iterated
Learning Model - which means, a generational model.
My Simulation
• What I judged the output on:– The original generation– The first generation
My Simulation
• What I measured:– Lexical retention– Lexical size– Hamming distance– Levenshtein distance– Phonotactic Development– Transitional Probability
Results
• Word Recognition: Pretty unsuccessful.
• Word Frequency: Wildly successful (100%)
• Transitional Probability: Alright.
• Transitional Counting: Better than alright, and each generation got better.
Results
Results
• Controls:– The WPT was very influential– Fixed original corpus: word recognition and
frequency did too well, while the transitional processes looked most like real language.
– Random original corpus: none of them did well.
Results
• Controls:– More words is better.– Shorter words is better.– Longer runs aren’t needed but are useful.
Disclaimers
• Online processing
• Memory constraints
• The WPT is unrealistic
• Words aren’t isolated
• Abstraction
• The digram analysis
Future Work?
• What about a Bayesian analysis? • How exactly would transition count and
probability be used in sequence?
• And anything you might raise now that shows that I need to redo this?
• thatisitiamdonenowthanks