learning bit by bit hidden markov models. weighted fsa weather the is outside 1.0.7.3

Learning Bit by Bit

Hidden Markov Models

Weighted FSA

weatherweatherTheThe isis

outsideoutside

1.0

.7

.3

Markov Chain

• Computing probability of an observed sequence of events

Markov Chain

weatherweather

TheThe

isis

outsideoutside

.7

.3

Observation = “The weather outside”

windwind

.5

.5

.1

.9

Parts of Speech

• Grammatical constructs like noun, verb

POS examples• N noun chair, bandwidth, pacing• V verb study, debate, munch• ADJ adjective purple, tall, ridiculous• ADV adverb unfortunately, slowly• P preposition of, by, to• PRO pronoun I, me, mine• DET determiner the, a, that, those

Parts of Speech-uses

• Speech recognition• Speech synthesis• Data mining• Translation

POS Tagging

• Words often have more than one POS: back– The back door = JJ– On my back = NN– Win the voters back = RB– Promised to back the bill = VB

• The POS tagging problem is to determine the POS tag for a particular instance of a word.

POS Tagging

• Sentence = sequence of observations• Ie. “Secretariat is expected to race tomorrow”

Disambiguating “race”

Hidden Markov Model

• Observed• Hidden

Hidden Markov Model

• 2 kinds of probabilities:– Tag transitions – Word likelihoods

Hidden Markov Model

• Tag transition prob = P( tag | previous tag)– ie. P(VB | TO)

Hidden Markov Model

• Word likelihood probability = P(word | tag)– ie. P(“race” | VB)

• Actual probabilities:– P (NN | TO) = .00047– P (VB | TO) = .83

• Actual probabilities:– P (NR| VB) = .0027– P (NR| NN) = .0012

• Actual probabilities:– P (race | NN) = .00057– P (race | VB) = .00012

Hidden Markov Model

• Probability “to race tomorrow” =“TO VB NR”• P(VB|TO) * P(NR|VB) * P(race|VB)• .83 * .0027 * .00012 = 0.00000026892

Hidden Markov Model

• Probability “to race tomorrow” =“TO NN NR”• P(NN|TO) * P(NR|NN) * P(race|NN)• .00047* .0012* .00057 = 0.00000000032148

Hidden Markov Model

• Probability “to race tomorrow” =“TO NN NR” = 0.00000000032148• Probability “to race tomorrow” =“TO VB NR”

= 0.00000026892

Bayesian Inference

• Correct answer = max (P (hypothesis | observed))

Bayesian Inference

• Prior probability = likelihood of the hypothesis

Bayesian Inference

• Likelihood = probability that the evidence matches the hypothesis

Bayesian Inference

• Bayesian vs. Frequentists• Subjectivity

Examples