System and method of call classification with context modeling based on composite words
A system and method of automatically classifying a communication involving at least one human, e.g., a human-to-human telephone conversation, into predefined categories of interest, e.g., “angry customer, etc. The system automatically, or semi-automatically with user interaction, expanding user input into semantically equivalent events. The system recognizes these events, each of which is given a confidence level, and classifies the telephone conversation based on an overall confidence level. In some embodiments, a different base unit is utilized. Instead of recognizing individual words, the system recognizes composite words, each of which is pre-programmed as an atomic unit, in a given context. The recognition includes semantically relevant composite words and contexts automatically generated by the system. The composite words based contextual recognition technique enables the system to efficiently and logically classifying and indexing large volumes of communications and audio collections such as call center calls, Webinars, live news feeds, etc.
1. A computer-implemented method of processing and classifying a communication involving at least one human, said method comprising:
processing by the computer said communication to detect one or more composite words in a given context C, wherein each composite word is a predefined sequence of words S pre-programmed as an atomic unit to prevent partial detection;
estimating a posterior probability P(S|C,X), wherein X is an acoustics vector sequence and wherein said posterior probability P(S|C,X)=P(X|C,S)P(S)/P(X|C,S)+P(X|C,S)), wherein S is an anti model of S;
determining an overall confidence level of events expressed in said one or more composite words occurring in said communication based on said posterior probability; and
classifying said communication into one or more categories based on detected events in accordance with said overall confidence level.
2. The method of claim 1 , further comprising:
approximating said anti model with acoustically similar sentences.
3. The method of claim 2 , further comprising:
performing a call classification on unrelated acoustic data;
obtaining false positives; and
configuring said anti model with said false positives.
4. The method of claim 1 , further comprising:
associating each composite word with a discrete or continuous state in accordance with a dynamic language model, wherein said dynamic language model specifies, given a point in said communication, prior probabilities for each composite word.
5. The method of claim 4 , wherein said dynamic language model affects said posterior probability based on categories found in said communication up to said point.
6. The method of claim 1 , further comprising:
enabling an end user to selectively review any of said events, said words or phrases, and said categories over a web browser application.
7. The method of claim 1 , in which said events are linguistic events.
8. The method of claim 1 , in which said events are non-linguistic events.
9. The method of claim 1 , further comprising:
assigning each event with a confidence level based on its occurrence.
10. The method of claim 1 , further comprising:
defining said events by examples, words, recordings, or user input.
11. The method of claim 1 , further comprising:
enabling a user to define an event by name, confidence level, speaker, start time, and end time.
12. The method of claim 1 , in which
said categories include one or more super-categories and one or more sub-categories.
13. The method of claim 1 , further comprising:
enabling a user to create a category and define it in terms of words, phrases, sentences, events, logical relationship to another category, or a combination thereof.
14. The method of claim 13 , further comprising:
automatically or semi-automatically expanding input from said user to generate computer output semantically equivalent to said user input.
15. The method of claim 14 , further comprising:
associating said user input and said semantically equivalent computer output with said category.
16. The method of claim 13 , further comprising:
enabling said user to provide or mark certain words as keywords belonging to said category, thereby instructing said computer to classify accordingly a communication containing one or more of said keywords.
17. The method of claim 16 , further comprising:
automatically generating additional words, phrases, sentences, or a combination thereof that are semantically equivalent to said keywords.
18. The method of claim 17 , further comprising:
consulting a knowledge base containing said additional words, phrases, sentences, or a combination thereof.
19. A digital computer system programmed to perform the method of claim 1 .
20. A computer readable medium tangibly embodying a computer-executable program of instructions implementing the method of claim 1 .
21. A system for processing and classifying a communication involving at least one human, said system comprising:
means for processing said communication to detect one or more composite words in a given context C, wherein each composite word is a predefined sequence of words S pre-programmed as an atomic unit to prevent partial detection;
means for estimating a posterior probability P(S|C,X), wherein X is an acoustics vector sequence and wherein said posterior probability P(S|C,X)=P(X|C,S)P(S)/P(X C,S)+P(X|C,S)), wherein S is an anti model of S;
means for determining an overall confidence level of events expressed in said one or more composite words occurring in said communication based on said posterior probability; and
means for classifying said communication into one or more categories based on detected events in accordance with said overall confidence level.
22. The system of claim 21 , further comprising:
means for approximating said anti model with acoustically similar sentences.
23. The system of claim 21 , further comprising:
means for performing a call classification on unrelated acoustic data;
means for obtaining false positives; and
means for configuring said anti model with said false positives.
24. The system of claim 21 , further comprising:
means for associating each composite word with a discrete or continuous state in accordance with a dynamic language model, wherein said dynamic language model specifies, given a point in said communication, prior probabilities for each composite word, wherein said dynamic language model affects said posterior probability based on categories found in said communication up to said point.
25. The system of claim 21 , further comprising:
means for enabling an end user to selectively review any of said events, said words or phrases, and said categories over a web browser application.
26. The system of claim 21 , in which
said events are linguistic events.
27. The system of claim 21 , in which
said events are non-linguistic events.
28. The system of claim 21 , further comprising:
means for assigning each event with a confidence level based on its occurrence.
29. The system of claim 21 , further comprising:
means for defining said events by examples, words, recordings, or user input.
30. The system of claim 21 , further comprising:
means for enabling a user to define an event by name, confidence level, speaker, start time, and end time.
31. The system of claim 21 , in which
said categories include one or more super-categories and one or more sub-categories.
32. The system of claim 21 , further comprising:
means for enabling a user to supply an original phrase with one or more words.
33. The system of claim 32 , further comprising:
means for analyzing said original phrase and identifying semantically important words thereof;
means for populating a list of semantically equivalent words for each semantically important word; and
means for determining whether each semantically equivalent word is statistically the same in context as said original phrase.
34. The system of claim 32 , further comprising:
means for automatically modifying the sentence structure of said original phrase to generate semantically equivalent phrases thereof.
35. The system of claim 32 , further comprising:
means for automatically adding words that precede or trail said original phrase to generate longer phrases that preserve the semantic meaning thereof.
36. The system of claim 32 , further comprising:
means for interacting with said user to learn about other phrases with which to interpret said original phrase; and
means for populating a list correlated with said original phrase to include said other phrases.
37. The system of claim 36 , further comprising:
means for generating follow-up questions on said original phrase; and
means for displaying said follow-up questions to said user.