Multi-task conditional random field models for sequence labeling
View Patent ↗Embodiments of a computer-implemented method for automatically analyzing a conversational sequence between multiple users are disclosed. The method includes receiving signals corresponding to a training dataset including multiple conversational sequences; extracting a feature from the training dataset based on predefined feature categories; formulating multiple tasks for being learned from the training dataset based on the extracted feature, each task related to a predefined label; and providing a model for each formulated task, the model including a set of parameters common to the tasks. The set includes an explicit parameter, which is explicitly shared with each of the formulated tasks. The method further includes optimizing a value of the explicit parameter to create an optimized model; creating a trained model for the formulated tasks using the optimized value of the explicit parameter; and assigning predefined labels for the formulated tasks to a live dataset based on the corresponding trained model.
1. A computer-implemented method for automatically analyzing a conversational sequence between a plurality of users, the method comprising:
receiving, using a data collection module on a computer comprising a processor and memory, signals corresponding to a training dataset including a plurality of data sequences related to the conversational sequence;
extracting, using a feature extraction module on the computer, at least one feature from the received training dataset based on predefined feature categories;
formulating, using a learning module on the computer, a plurality of tasks for being learned from the training dataset based on the extracted at least one feature, wherein each of the plurality of tasks is related to at least one predefined label;
providing, using the learning module on the computer, a model for each of the plurality of formulated tasks, wherein the model includes one or more parameters having a set of parameters common to the plurality of formulated tasks, wherein the set of parameters includes at least one label dependency factor that is an explicit parameter being explicitly shared with each of the plurality of formulated tasks;
optimizing, using the learning module on the computer, values for the one or more parameters and the at least one explicit parameter to create an optimized model; creating, using the learning module on the computer, a trained model for each of the plurality of formulated tasks using an optimized value of the at least one explicit parameter and corresponding values of the one or more parameters;
assigning, using a classification module on the computer, the at least one predefined label for each of the plurality of formulated tasks on to a live dataset based on the corresponding created trained model; and
outputting, using the computer, signals corresponding to the live dataset assigned with the at least one predefined label for each of the plurality of formulated tasks.
2. The method of claim 1 , wherein the step of creating a trained model further comprises:
evaluating, using the learning module, the optimized model based on conventional models determined for a test dataset;
tuning, using the learning module, the optimized value for the at least one explicit parameter based on the evaluated optimized model; and
creating, using the learning module, a trained model for each of the plurality of formulated tasks using the tuned value of the at least one explicit parameter and corresponding values of the one or more parameters.
3. The method of claim 2 , wherein the optimized value for the at least one explicit parameter is tuned until a conditional log likelihood of a corresponding task among the plurality of formulated tasks is maximized, wherein the optimized value is tuned by varying the value of a predetermined weighted factor to optimize the corresponding one or more parameters for at least one task among the plurality of formulated tasks.
4. The method of claim 1 , wherein the predefined feature categories include at least one of a “Word 1-grams and 2-grams”, “Segment position in the conversation”, “Segment position in a conversation sequence”, “Sender”, “Contains email”, “#Upper case”, “# Punctuation”, “#Special punctuation”, “Positive Sentiment”, “Negative Sentiment”, “Category of previous sequence”, “Category of previous sequence of the same author”, and “Category of previous sequence of a different author”.
5. The method of claim 1 , wherein a first task among the plurality of formulated tasks is associated with a first predefined label and a second task among the plurality of formulated tasks is associated with a second predefined label, wherein the first predefined label and the second predefined label are correlated.
6. The method of claim 5 , wherein the first task corresponds to a task for determining an issue status in the conversational sequence and a second task corresponds to a task for determining a nature of the conversational sequence.
7. The method of claim 1 , wherein the plurality of data sequences are natural language text sequences including a plurality of sentences, wherein each of the plurality of sentences includes one or more words.
8. The method of claim 1 , wherein the trained model has a predefined accuracy, which is inversely proportional to the complexity of the training dataset, wherein the complexity of the training dataset corresponds to at least one of size of the training dataset and a number of pre-labeled data sequences in the training dataset.
9. The method of claim 1 , wherein the plurality of data sequences are prelabeled using a plurality of predetermined labels corresponding to at least one of the plurality of formulated tasks.
10. The method of claim 9 , wherein the plurality of predetermined labels includes at least one of “OPEN”, “SOLVED”, “CLOSED”, and “CHANGED CHANNEL” corresponding to a task for determining an issue status in the conversational sequence among the plurality of formulated tasks.
11. The method of claim 9 , wherein the plurality of predetermined labels includes at least one of “COMPLAINT”, “REQUEST”, “APOLOGY”, “ANSWER”, “RECEIPT”, “COMPLIMENT”, “RESPONSE TO POSITIVITY”, “GREETING”, “THANK”, “ANNOUNCEMENT”, “SOLVED”, and “OTHER” corresponding to a task for determining
a nature of the conversational sequence among the plurality of formulated tasks.
12. The method of claim 1 , further comprising, determining, by the computer using the trained model, a status of an issue in the conversational sequence.
13. The method of claim 1 , further comprising, determining, by the computer using the trained model, a nature of the conversational sequence.
14. The method of claim 1 , wherein the set of parameters include at least one parameter having variance specific to each of the plurality of formulated tasks.
15. A system for automatically analyzing a conversation between a plurality ofusers, the system comprising:
a data collection module on a computer comprising a processor and memory configured to receive signals corresponding to a training dataset including a plurality of data sequences related to the conversation;
a feature extraction module on the computer configured to extract at least one feature from the received training dataset based on predefined feature categories;
a learning module on the computer configured to:
formulate a plurality of tasks for being learned from the training dataset based on the extracted at least one feature, wherein each of the plurality of tasks is related to at least one predefined label;
provide a model for each of the plurality of formulated tasks, wherein the model includes one or more parameters having a set of parameters common to the plurality of formulated tasks, wherein the set of parameters includes at least one label dependency factor, that is an explicit parameter, which is explicitly shared with each of the plurality of formulated tasks;
optimize values for the one or more parameters and the at least one explicit parameter to create an optimized model; and
create a trained model for each of the plurality of formulated tasks using an optimized value of the at least one explicit parameter and corresponding values of the one or more parameters; and
a classification module on the computer configured to assign the at least one predefined label for each of the plurality of formulated tasks on to a live dataset based on the created trained model, wherein the computer is configured to output signals corresponding to the live dataset assigned with the at least one predefined label for each of the plurality of formulated tasks.
16. The system of claim 15 , wherein the set of parameters include at least one parameter having variance specific to each of the plurality of formulated tasks.
17. The system of claim 15 , wherein the learning module is configured to vary the value of a predetermined weighted factor for tuning the value of the at least one explicit parameter for optimization until a conditional log likelihood of a corresponding task among the plurality of formulated tasks is maximized.
18. The system of claim 15 , wherein the predefined feature categories include at least one of a “Word 1-grams and 2-grams”, “Segment position in the conversation”, “Segment position in a conversation sequence”, “Sender”, “Contains email”, “#Upper case”, “# Punctuation”, “#Special punctuation”, “Positive Sentiment”, “Negative Sentiment”, “Category of previous sequence”, “Category of previous sequence of the same author”, and “Category of previous sequence of a different author”.
19. The system of claim 15 , wherein a first task among the plurality of formulated tasks is associated with a first predefined label and a second task among the plurality of formulated tasks is associated with a second predefined label, wherein the first predefined label and the second predefined label are correlated.
20. The system of claim 19 , wherein the first task corresponds to a task for determining an issue status in the conversation and a second task corresponds to a task for determining a nature of the conversation.
21. The system of claim 15 , wherein the plurality of data sequences are natural language text sequences including a plurality of sentences, wherein each of the plurality of sentences include one or more words.
22. The system of claim 15 , wherein the plurality of data sequences are prelabeled using a plurality of predetermined labels corresponding to at least one of the plurality of formulated tasks.
23. The system of claim 22 , wherein the plurality of predetermined labels includes at least one of “OPEN”, “SOLVED”, “CLOSED”, and “CHANGED CHANNEL” corresponding to a task for determining an issue status in the conversation among the plurality of formulated tasks.
24. The system of claim 22 , wherein the plurality of predetermined labels includes at least one of “COMPLAINT”, “REQUEST”, “APOLOGY”, “ANSWER”, “RECEIPT”, “COMPLIMENT”, “RESPONSE TO POSITIVITY”, “GREETING”, “THANK”, “ANNOUNCEMENT”, “SOLVED”, and “OTHER” corresponding to a task for determining
a nature of the conversation among the plurality of formulated tasks.
25. The system of claim 15 , wherein the trained model has a predefined accuracy, which is inversely proportional to the complexity of the training dataset, wherein the complexity of the training dataset corresponds to at least one of size of the training dataset and a number of pre-labeled data sequences in the training dataset.