Automatic classification of phone calls using representation learning based on the hierarchical pitman-yor process
Embodiments of the disclosed technology include a representation learning model for classification of natural language text. In embodiments, a classification model comprises a feature model and a classifier. The feature model may be hierarchical in nature: data may pass through a series of representations, decreasing in specificity and increasing in generality. Intermediate levels of representation may then be used as automatically learned features to train a statistical classifier. Specifically, the feature model may be based on a hierarchical Pitman-Yor process. In embodiments, once the feature model has been expressed as a Bayesian Belief Network and some aspect of the feature model has been selected for prediction, the feature model may be attached to the classifier. In embodiments, after training, potentially using a mix of labeled and unlabeled data, the classification model can be used to classify documents such as call transcripts based on topics of conversation represented in the transcripts.
1. A computer-implemented method, comprising:
digitally storing first call transcript data that is associated with an observed label of a set of observed labels, the first call transcript data comprising an electronic digital representation of a verbal transcription of a call between a first person of a first person type and a second person of a second person type, the first call transcript data having been created based on speech-to-text recognition of an audio recording of the call, the first call transcript data being divided into first person type data comprising words spoken by the first person in the call and second person type data comprising words spoken by the second person in the call;
digitally generating and storing a machine learning statistical topic model in computer memory, the topic model comprising a word branch, a topic branch, and a classifier that defines a joint probability distribution over topic vectors and observed labels, the classifier being conjoined to the topic branch, the topic model simultaneously modeling the first person type data as a function of a first probability distribution of words used by the first person type for one or more topics and the second person type data as a function of a second probability distribution of words used by the second person type for the one or more topics, both the first probability distribution of words and the second probability distribution of words being modeled as a function of a third probability distribution of words for the one or more topics;
wherein the classifier is a linear classifier that has been trained to use a determined probability distribution of topics and a determined topic importance to determine classifications of target calls;
programmatically training, using a set of call transcript data that includes the first call transcript data, the topic model using the classifier, the set of call transcript data comprising at least one call transcript data not associated with an observed label;
receiving target call transcript data comprising an electronic digital representation of a verbal transcription of a target call;
programmatically determining, using the topic model and the linear classifier, at least one of one or more topics of the target call and a classification of the target call;
digitally storing the target call transcript data with additional data indicating at least one of the one or more topics of the target call or the classification of the target call;
wherein the word branch and the topic branch are conditionally independent;
wherein the topic model uses probability distributions to infer any data missing from the first call transcript data.
2. The method of claim 1 , the word branch of the topic model being configured to model words as parts of phrases and the topic branch of the topic model being configured to model transition probabilities among topics.
3. The method of claim 1 , the topic model additionally modeling the first call transcript data as a function of a plurality of topics which are modeled as a function of a probability distribution of topics.
4. The method of claim 3 , the probability distribution of topics being modeled as a function of an inferred prior probability distribution which is modeled as a function of a flat prior distribution.
5. The method of claim 4 , the probability distribution of topics being modeled as a function of the inferred prior probability distribution using a Pitman-Yor Process and the inferred prior probability distribution being modeled as a function of the flat prior distribution using a Pitman-Yor Process.
6. The method of claim 1 , the third probability distribution of words for each topic being modeled as a function of an inferred prior probability distribution which is modeled as a function of a flat prior distribution.
7. The method of claim 6 , the third probability distribution of words for each topic being modeled as a function of the inferred prior probability distribution using a Pitman-Yor Process and the inferred prior probability distribution being modeled as a function of the flat prior distribution using a Pitman-Yor Process.
8. The method of claim 1 , determining the one or more topics of the call using the topic model by inverting a Bayesian Belief Network using Bayes' theorem.
9. The method of claim 1 , the linear classifier comprising one of a Finnish Horseshoe Model, an L2 Logistic Regression, or a Logistic Regression using a Horseshoe Potential.
10. The method of claim 9 , the linear classifier being a Finnish Horseshoe Model and wherein a t parameter of the Finnish Horseshoe Model is inferred.
11. The method of claim 1 , training the topic model involving a Rao-Blackwellization process and using a Hamiltonian Monte Carlo algorithm to update the classifier.
12. The method of claim 1 , the first person type being a caller type and the second person type being an agent type.
13. The method of claim 1 , further comprising providing, to a client computing device, topic information indicating, for each of a plurality of topics, a number or percentage of calls received for that topic over a particular period of time.
14. One or more non-transitory storage media storing instructions which, when executed by one or more processors, cause the one or more processors to execute:
digitally storing first call transcript data that is associated with an observed label of a set of observed labels, the first call transcript data comprising an electronic digital representation of a verbal transcription of a call between a first person of a first person type and a second person of a second person type, the first call transcript data having been created based on speech-to-text recognition of an audio recording of the call, the first call transcript data being divided into first person type data comprising words spoken by the first person in the call and second person type data comprising words spoken by the second person in the call;
digitally generating and storing a machine learning statistical topic model in computer memory, the topic model comprising a word branch, a topic branch, and a classifier that defines a joint probability distribution over topic vectors and observed labels, the classifier being conjoined to the topic branch, the topic model simultaneously modeling the first person type data as a function of a first probability distribution of words used by the first person type for one or more topics and the second person type data as a function of a second probability distribution of words used by the second person type for the one or more topics, both the first probability distribution of words and the second probability distribution of words being modeled as a function of a third probability distribution of words for the one or more topics;
wherein the classifier is a linear classifier that has been trained to use a determined probability distribution of topics and a determined topic importance to determine classifications of target calls;
programmatically training, using a set of call transcript data that includes the first call transcript data, the topic model using the classifier, the set of call transcript data comprising at least one call transcript data not associated with an observed label;
receiving target call transcript data comprising an electronic digital representation of a verbal transcription of a target call;
programmatically determining, using the topic model and the linear classifier, at least one of one or more topics of the target call and a classification of the target call;
digitally storing the target call transcript data with additional data indicating at least one of the one or more topics of the target call or the classification of the target call;
wherein the word branch and the topic branch are conditionally independent;
wherein the topic model uses probability distributions to infer any data missing from the first call transcript data.
15. The media of claim 14 , further comprising sequences of instructions which, when executed using the one or more processors, cause the one or more processors to execute determining the one or more topics of the call using the topic model by inverting the topic model using a Bayesian Belief Network.
16. The media of claim 14 , the linear classifier comprising one of a Finnish Horseshoe Model, an L2 Logistic Regression, or a Logistic Regression using a Horseshoe Potential.
17. The media of claim 14 , further comprising sequences of instructions which, when executed using the one or more processors, cause the one or more processors to execute training the topic model using a Rao-Blackwellization process and a Hamiltonian Monte Carlo algorithm to update the classifier.
18. The media of claim 14 , the first person type being a caller type and the second person type being an agent type.
19. The media of claim 14 , further comprising sequences of instructions which, when executed using the one or more processors, cause the one or more processors to execute providing, to a client computing device, topic information indicating, for each of a plurality of topics, a number or percentage of calls received for that topic over a particular period of time.
20. A system comprising:
one or more processors;
a memory storing instructions which, when executed by the one or more processors, causes the one or more processors to execute:
digitally storing first call transcript data that is associated with an observed label of a set of observed labels, the first call transcript data comprising an electronic digital representation of a verbal transcription of a call between a first person of a first person type and a second person of a second person type, the first call transcript data having been created based on speech-to-text recognition of an audio recording of the call, the first call transcript data being divided into first person type data comprising words spoken by the first person in the call and second person type data comprising words spoken by the second person in the call;
digitally generating and storing a machine learning statistical topic model in computer memory, the topic model comprising a word branch, a topic branch, and a classifier that defines a joint probability distribution over topic vectors and observed labels, the classifier being conjoined to the topic branch, the topic model simultaneously modeling the first person type data as a function of a first probability distribution of words used by the first person type for one or more topics and the second person type data as a function of a second probability distribution of words used by the second person type for the one or more topics, both the first probability distribution of words and the second probability distribution of words being modeled as a function of a third probability distribution of words for the one or more topics;
wherein the classifier is a linear classifier that has been trained to use a determined probability distribution of topics and a determined topic importance to determine classifications of target calls;
programmatically training, using a set of call transcript data that includes the first call transcript data, the topic model using the classifier, the set of call transcript data comprising at least one call transcript data not associated with an observed label;
receiving target call transcript data comprising an electronic digital representation of a verbal transcription of a target call;
programmatically determining, using the topic model and the linear classifier, at least one of one or more topics of the target call and a classification of the target call;
digitally storing the target call transcript data with additional data indicating at least one of the one or more topics of the target call or the classification of the target call;
wherein the word branch and the topic branch are conditionally independent;
wherein the topic model uses probability distributions to infer any data missing from the first call transcript data.