online_em.m

隐马尔科夫模型对文本信息进行抽取利用MATLAB实现

function [prior, transmat, obsmat, exp_num_visits, exp_num_trans, exp_num_emit, gamma, ll] = online_em(...
    prior, transmat, obsmat, exp_num_visits, exp_num_trans, exp_num_emit, decay, dirichlet, data, ...
    act, adj_prior, adj_trans, adj_obs)
% ONLINE_EM Adjust the parameters using a weighted combination of the old and new expected statistics
%
% [prior, transmat, obsmat, exp_num_visits, exp_num_trans, exp_num_emit, gamma, ll] = online_em(...
%    prior, transmat, obsmat, exp_num_visits, exp_num_trans, exp_num_emit, decay, dirichlet, data, ...
%    act, adj_prior, adj_trans, adj_obs)
% 
% 0 < decay < 1, with smaller values meaning the past is forgotten more quickly.
% (We need to decay the old ess, since they were based on out-of-date parameters.)
% The other params are as in learn_hmm.
% We do a single forwards-backwards pass on the provided data.

if ~exist('act'), act = []; end
if ~exist('adj_prior'), adj_prior = 1; end
if ~exist('adj_trans'), adj_trans = 1; end
if ~exist('adj_obs'), adj_obs = 1; end

% E step
olikseq = mk_dhmm_obs_lik(data, obsmat);
if isempty(act)
  [alpha, beta, gamma, xi, ll] = forwards_backwards(prior, transmat, olikseq);
else
  [alpha, beta, gamma, xi, ll] = forwards_backwards(prior, transmat, olikseq, [], [], act);
end

% Increment ESS
exp_num_visits = decay*exp_num_visits + sum(gamma, 2);

[S O] = size(obsmat);
exp_num_emit = decay*exp_num_emit + dirichlet*ones(S,O);
T = length(data);
if T < O
  for t=1:T
    o = data(t);
    exp_num_emit(:,o) = exp_num_emit(:,o) + gamma(:,t);
  end
else
  for o=1:O
    ndx = find(data==o);
    if ~isempty(ndx)
      exp_num_emit(:,o) = exp_num_emit(:,o) + sum(gamma(:, ndx), 2);
    end
  end
end

if T > 1
  if isempty(act)
    exp_num_trans = decay*exp_num_trans + sum(xi,3);
  else
    % act(2) determines Q(2), xi(:,:,1) holds P(Q(1), Q(2))
    A = length(transmat);
    for a=1:A
      ndx = find(act(2:end)==a);
      if ~isempty(ndx)
	exp_num_trans{a} = decay*exp_num_trans{a} + sum(xi(:,:,ndx), 3);
      end
    end
  end
end


% M step

% If a param is non-adjustable, we needn't have computed its ess
if adj_prior
  prior = normalise(exp_num_visits);
end
if adj_obs
  obsmat = mk_stochastic(exp_num_emit);
end
if adj_trans & (T>1)
  if isempty(act)
    transmat = mk_stochastic(exp_num_trans);
  else
    for a=1:A
      transmat{a} = mk_stochastic(exp_num_trans{a});
    end
  end
end