Discriminative training of language models for text and speech classification
Methods are disclosed for estimating language models such that the conditional likelihood of a class given a word string, which is very well correlated with classification accuracy, is maximized. The methods comprise tuning statistical language model parameters jointly for all classes such that a classifier discriminates between the correct class and the incorrect ones for a given training sentence or utterance. Specific embodiments of the present invention pertain to implementation of the rational function growth transform in the context of a discriminative training technique for n-gram classifiers.
1. A computer-implemented classification method, comprising:
estimating a set of parameters for each of a plurality of n-gram language models, each n-gram language model being associated with a different category of subject matter, the sets of parameters being estimated utilizing training data that includes main data and held-out data, the sets of parameters including relative frequencies that are estimated using the main data, the relative frequencies being estimated utilizing a context sensitive conditional maximum likelihood (CML) weight, the context sensitive CML weight being set to a maximum value that is greater than zero at each iteration and then lowered separately for each context, the relative frequencies being iteratively re-estimated utilizing a ratio of a maximum likelihood count to a CML count;
utilizing a computer processor that is a functional component of a computer to produce the sets of parameters for at least two of the n-gram language models jointly in relation to one another through an application of a rational function growth transform, a number of iterations of the rational function growth transform being tuned using the held-out data, the maximum value of the CML weight also being tuned using the held-out data; and
utilizing the sets of parameters of the plurality of n-gram language models that are produced at least partially jointly in relation to one another as a basis for supporting a determination as to which of the categories of subject matter is best correlated to a given natural language input, wherein the category of subject matter that is best correlated to the given natural language input is information other than a textual representation of the given natural language input itself.
2. The method of claim 1 , wherein
the category of subject matter that is best correlated to the given natural language input includes a plurality of categories, the plurality of categories being ranked and displayed to a user through a user interface such that the user can select one of the plurality of categories, and wherein producing the sets for at least two of the n-gram language models jointly in relation to one another comprises producing all sets of parameters of the plurality of n-gram language models jointly in relation to one another.
3. The method of claim 1 , wherein
each of the plurality of n-gram language models is associated with a class, wherein producing the sets of parameters for at least two of the n-gram language models jointly comprises producing at least two of the sets of parameters such that a first particular class will have a higher probability value associating it to a particular word string, and such that a second class will have a lower probability value associating it to the particular word string, and wherein the determination as to which of the categories of subject matter is best correlated to a given natural language input comprises outputting a list of results.
4. The method of claim 3 , wherein
the particular word string is derived from a training sentence, and wherein the training sentence is preprocessed to remove certain elements to form the particular word string.
5. The method of claim 3 , wherein the particular word string is derived from a speech utterance, and wherein acoustic confusability between words is accounted for.
6. The method of claim 1 , wherein producing the sets of parameters for at least two of the n-gram language models jointly comprises producing multiple sets of parameters jointly such that the n-gram language models will discriminate between a correct class and incorrect classes for a given training input, and wherein the natural language input comprises a vector representation.
7. The method of claim 1 , further comprising training the sets of parameters to accommodate unseen data based upon a recursive deleted interpolation scheme.
8. A computer-implemented classification method, comprising:
estimating a set of parameters for each of a plurality of n-gram language models, each n-gram language model being associated with a different category of subject matter, the sets of parameters being estimated utilizing training data that includes main data and held-out data, the sets of parameters including relative frequencies that are estimated using the main data, the relative frequencies being estimated utilizing multiple iterations of a rational function growth transformation (RFGT) and a context sensitive conditional maximum likelihood (CML) weight, the context sensitive CML weight being set to a maximum value that is greater than zero at each iteration and then lowered separately for each context, the multiple iterations of the RFGT being preset to a number N, the maximum value of the CML weight being determined utilizing a fix grid having a range of values and steps;
utilizing a computer processor that is a functional component of a computer to produce the sets of parameters for at least two of the n-gram language models jointly in relation to one another through an application of a rational function growth transform, a number of iterations of the rational function growth transform being tuned using the held-out data, the maximum value of the CML weight also being tuned using the held-out data; and
utilizing the sets of parameters of the plurality of n-gram language models that are produced at least partially jointly in relation to one another as a basis for supporting a determination as to which of the categories of subject matter is best correlated to a given natural language input, wherein the category of subject matter that is best correlated to the given natural language input is information other than a textual representation of the given natural language input itself.
9. The method of claim 8 , wherein for each value of the context sensitive CML weight no more than the number N iterations are performed, and a conditional likelihood of the main data increases at each of the N iterations.
10. The method of claim 8 , and further comprising:
pooling the main data and the held-out data; and
re-training the plurality of n-gram language models utilizing the pooled data.