Method for recognizing statistical voice language
The present invention is a method for recognizing a statistical voice language using a statistical technique without using a manually tagged corpus. The method comprises: a dialog act clustering step of clustering speech utterances of sentences based on similar dialog acts; a named entity clustering step of extracting a named entity candidate group from the result of the dialog act clustering step and clustering named entities based on the neighboring contextual information of the extracted named entity candidate group; and a main act clustering step of clustering main acts for each region based on the clustered dialog acts and named entity.
1. A computerized statistical method for speech language understanding, performed by a processor in a computing apparatus, sequentially comprising:
dialog act clustering including clustering spoken sentences based on similar dialog acts included in the spoken sentences;
named entity clustering including extracting a group of named entity candidates from a result of the dialog act clustering, and clustering named entities based on context information around the extracted group of candidate named entities; and
main act clustering including clustering main acts for each domain based on a result of the named entity clustering,
wherein the named entity clustering extracts a group of candidate named entities using sources of words included in the spoken sentences,
wherein when the group of candidate named entities includes a plurality of consecutive words, the plurality of consecutive words are segmented using a stickiness function.
2. The method of claim 1 , wherein the method is performed based on a nonparametric and unsupervised learning method, without using contrived manual tagging.
3. The method of claim 1 , wherein the dialog act clustering is performed using a non-parametric Bayesian hidden Markov model, under conditions that the dialog act is a hidden state and words included in the spoken sentences are observation values.
4. The method of claim 3 , wherein the non-parametric Bayesian hidden Markov model is a hierarchical Dirichlet process hidden Markov model (HDP-HMM).
5. The method of claim 4 , wherein a transition probability of the dialog act is determined by a Dirichlet process.
6. The method of claim 3 , wherein in the dialog act clustering, the dialog act clustering is performed using sources of the words included in the spoken sentences, under conditions that the dialog act is a hidden state and the words are observation values.
7. The method of claim 3 , wherein the dialog act clustering finally infers the dialog act using Gibbs sampling.
8. The method of claim 1 , further comprising classifying domains of the spoken sentences using a distribution of domain-specific words that is obtained in the dialog act clustering.
9. The method of claim 1 , wherein the stickiness function is a point mutual information (PMI) function.
10. The method of claim 1 , wherein the context information around the extracted group of candidate named entities is obtained by applying a hierarchical Dirichlet process algorithm to a predetermined number of words positioned before and after the extracted group of the named entity candidates.
11. The method of claim 1 , wherein the main act clustering is performed using a non-parametric Bayesian hidden Markov model, under conditions that the main act is a hidden state and words included in the spoken sentences, named entities extracted from the spoken sentences and system activities corresponding to the spoken sentences are observation values.
12. The method of claim 11 , wherein the non-parametric Bayesian hidden Markov model is a hierarchical Dirichlet process hidden Markov model.
13. The method of claim 12 , wherein a transition probability of the main act is determined by a Dirichlet process.
14. The method of claim 13 , further comprising generating an agenda graph using main acts derived from the main act clustering and the transition probability.