More Segmentation (6)

are transitional probabilities a solution for word segmentation

no

how do we evaluate a segmentation algorithm and what does it mean

it would work “well” - high-accuracy and human-like

what are true positives in the context of word segmentation

a segmentation was made where there should be one (correct)

what are false positives in the context of word segmentation

a segmentation was made were there shouldn’t be one (incorrect)

what are true negative in the context of word segmentation

a segmentation wasn’t made where there shouldn’t be one (correct)

what are false negative in the context of word segmentation

a segmentation wasn’t made were there should be one (incorrect)

what is “percision” in an algorithm

of all the things it chooses, how many where correct

what is “recall” in algorithms, what is a flaw in it

how many of the things did it get, error is it can take everything to get 100% recall 

what is an “f-score” in algorithms

a combonation measures of percision and recall

what is the equation for percision

P = TP / (TP+FP)

what is the equation for recall

R = TP / (TP+FN)

what is the equation for an F-score

F = 2 x [(PxR) / (P+R)]

what is gold and prediction

gold - what’s true

prediciton - our guess

what is a baseline in thinking ab segmentation

if we did something obvious and dumb, how well would it do? 

if we said every syll in eng was a boundary, how well would we do? is this helpful compared to CHILDES?

87.4% correct, 1.2 syll per word in CDS of CHILDES

what two things make english segmentation easier and why

high syll diversity (makes TransProb calculations cleaner)

words unlikely to be substrings of other words

what 2 things make langauge segmentation more difficult 

high-average syll in words
low syll diversity 

what forms edges within an utterance

prosodic boundaries

why do edges help w word segmentation

if the word is at an edge, you get a boundary for free

what has been experimentally showed about edges and why are they important

children are sensitive to edges, more important than TP

what is shown about edges algorithmically

they’re nessessary

what are the 2 parts of distibutional learning

batch + statistical/distibutional

what is batch in distibutional learning 

processing a lot of data at once 

what is statistical/distibutional in distibutional learning

they rely on probabilities to do most of the work

what are the two parts of mechanistic learning

online/incremental + mechanistic/algorithmic

what is online/incremental in mechanistic learning 

process data bit by bit as it comes in 

what is mechanistic/algorithmic in mechanistic learning

focuses on the process rather than the stats

what is the basic idea behind subtrastic segmentors and by when is it shown children start this

0;6

children keep track of words they already know (lexicon) and subtract them from new utts

describe the 3 steps of the subtractive algorithm 

(if utt starts w known word)

segment it off

increment the score of that word (repeat on remainder)

add what’s left of utt to lexicon 

can you add more rules to a subtractive algorithm for better results

yes

what 2 things are TP and what 2 things are subtractive

TP: batch + distibutional
Subtractive: online + mechanistic 

what are the cognitive implications of TP and subtractive methods

TP - chidlren track words and probabiliteis of B|A syll bigram

Subtractive - children count and track words

don’t confused consistancy with…

explanation