Dept. of Linguistics, Indiana University Fall 2009

Transcription

1 1 / 14 Markov L645 Dept. of Linguistics, Indiana University Fall 2009

2 2 / 14 Markov (1) (review) Markov A Markov Model consists of: a finite set of statesω={s 1,...,s n }; an signal alphabetσ={σ 1,...,σ m }; an n n state transition matrix P=[p ij ] where p ij = P(ξ t+1 = s j ξ t = s i ); an n m signal matrix A=[a ij ], which for each state-signal pair determines the probability a ij = p(η t =σ j ξ t = s i ) that signalσ j will be emitted given that the current state is s i ; and an initial vector v=[v 1,...,v n ] where v i = P(ξ 1 = s i ).

3 3 / 14 HMM Example 1: Crazy Softdrink Machine 0.5 Markov 0.7 ColaPref. 0.3 Iced Tea Pref. 0.5 with emission probabilities: cola ice-t lem CP IP from: Manning/Schütze; p. 321

4 4 / 14 Markov (2) (Review) Markov p (t) (s i,σ j )=p (t) (s i ) p(η t =σ j ξ t = s i ) where p (t) (s i ) is the ith element of the vector vp t 1. The probability that signalσ j will be emitted at time t is then: p (t) (σ j )= n p (t) (s i,σ j )= i=1 n p (t) (s i ) p(η t =σ j ξ t = s i ) i=1 Thus if p (t) (σ j ) is the probability of the model emitting signalσ j at time t, i.e., after t 1 steps, then [p (t) (σ 1 ),...,p (t) (σ m )]=vp t 1 A

5 5 / 14 (1) Markov Let O Σ be a known sequence of observed signals and let S Ω be the sequence of states in which O is emitted. If it is not possible to observe the sequence of states S 1,...,S T of a Markov model, but only the sequence n 1,...,n T, the model is called a hidden Markov model (an HMM). (Krenn/Samuelsson, p. 43) In an HMM, you don t know the state sequence that the model passes through, but only some probabilistic function of it. (Manning/Schütze, p. 320) Our best guess at S is the sequence maximizing max P(S O) S

6 6 / 14 (2) Markov Prototypical tasks to which hidden Markov models are applied include the following. Given a sequence of signals O=(σ i1,...,σ it ): Estimate the probability of observing this particular signal sequence. e.g., language identification Determine the most probable state sequence that can give rise to this signal sequence. e.g., POS tagging & speech recognition Determine the set of model parametersλ=(p, A, v) maximizing the probability of this signal sequence.

7 7 / 14 (3) Markov 2 types of representations: state emission HMM the symbol emitted at time t depends only on the state at time t arc emission HMM the symbol emitted at time t depends both on states at time t and time t+ 1 Manning/Schütze do arc emission, Krenn/Samuelsson do state emission we ll do state emission

8 8 / 14 HMM Application 1: POS tagging Markov The set of observable signals are the words of an input text. The states are the set of tags that are to be assigned to the words of input text. The task consists in finding the most probable sequence of states that explains the observed words. This will assign a particular state to each signal, i.e., a tag to each word.

9 Example HMM Assume that we have DET, N, and VB as our hidden states, and we have the following transition matrix (A):... emission matrix (B): DET N VB DET N VB Markov dogs bit the chased a these cats... DET N VB and initial probability matrix (π): DET 0.7 N 0.2 VB / 14

10 10 / 14 State sequences Markov If we generate the bit dogs, we don t know which tag sequence generated it: DET N VB? DET N N? DET VB N? DET VB VB? Each has different probabilities, so we need an algorithm which will give us the best sequence of states (i.e., tags) for our given sequence of words

11 11 / 14 HMM Application 2: Speech recognition Markov The set of observable signals are (some representation of the) acoustic signals. The states are the possible words that these signals could arise from. The task consists in finding the most probable sequence of words that explains the observed acoustic signals. This is a slightly more complicated situation, since the acoustic signals do not stand in a one-to-one correspondence with the words.

12 12 / 14 Three Fundamental Problems for HMMs (1) Markov Calculating the Probability of an Observation Sequence: Given the observation sequence O= O 1 O 2...O T, and a modelλ=< P, A, v>, how to (efficiently) compute the probability of the observation sequence, given the model: P(O λ) Method of Choice: Trellis algorithm using forward or backward accumulator variables.

13 13 / 14 Three Fundamental Problems for HMMs (2) Markov Finding the Optimal State Sequence: Given the observation sequence O= O 1 O 2...O T and the modelλ, how do we choose the most likely state sequence that corresponds to O : max S P(S O) Method of Choice: Viterbi-style dynamic programming algorithm using a trellis and backpointers

14 14 / 14 Three Fundamental Problems for HMMs (3) Markov Parameter Estimation: How to estimate the model parameters λ=< P, A, v> to maximize P(O λ) Method of Choice: Baum-Welch algorithm using forward-backward re-estimation