Detection of communication topic change

Patent metadata

Field	Value
Publication number	US-9645703-B2
Application number	US-201414277290-A
Country	US
Kind code	B2
Filing date	May 14, 2014
Priority date	May 14, 2014
Publication date	May 9, 2017
Grant date	May 9, 2017

Abstract

A computer processor determines a first span of a communication, wherein a span includes content associated with one or more dialog statements. If the content of the first span contains one or more topic change indicators which are identified by at least one detector of a learning model, the computer processor, in response, generates scores for each of the one or more indicators. The computer processor aggregates scores of the one or more indicators of the first span, which may be weighted, to produce an aggregate score. The computer processor compares the aggregate score to a threshold value, wherein the threshold value is determined during training of the learning model, and the computer processor, in response to the aggregate score crossing the threshold value, determines a topic change has occurred within the first span.

First claim

What is claimed is:

1. A method for determining a topic change of a communication, the method comprising:
determining, by a computer processor, a communication containing a set of dialog statements, wherein a first span of the communication includes one or more dialog statements of the set of dialog statements;
determining, by the computer processor, if the one or more dialog statements of the first span include one or more indicators of a topic change, wherein the one or more indicators are identified by at least one detector of a learning model;
responsive to determining the first span includes the one or more indicators of the topic change, generating, by the computer processor, a score for the one or more indicators, based on the learning model;
responsive to the score for the one or more indicators triggering a threshold condition, determining, by the computer processor, a topic change within the first span,
wherein the threshold condition determined during training of the learning model, includes:
determining, by the computer processor, a weighted value for the at least one detector, based on heuristics;
receiving input of labelled communication dialog statements, wherein the labelled communication dialog statements include one or more topic change indicators that are known, the one or more topic change indicators corresponding to the at least one detector;
adjusting, by the computer processor, the weighted value of the at least one detector in response to a delta between an output of scores of the at least one detector of the learning model and scores of the one or more topic change indicators that are known; and
determining, by the computer processor, the threshold condition in response to achieving an acceptable minimum for the delta between the output of the scores which are determined by the at least one detector of the learning model and the scores of the one or more topic change indicators that are known.

2. The method of claim 1 , further comprising:
receiving, by a computer processor, the communication including the first span;
generating, by the computer processor, a second span based on adjusting boundaries of the first span by performing at least one of, adding to the first span, one or more dialog statements of the set of dialog statements not included in the first span, and removing one or more dialog statements from the first span;
determining, by the computer processor, a score for the first span and a score for the second span, wherein the score for the first span and the score for the second span is based on a topic of the first span and a topic of the second span, respectively;
responsive to the score of the second span being more favorable than the score of the first span, extracting, by the computer processor, one or more features from the one or more dialog statements of the second span not included in the first span; and
training, by the computer processor, the learning model to determine a topic change, based, at least in part, on including the one or more features from the one or more dialog statements of the second span, in at least one detector of the learning model.

3. The method of claim 2 , wherein extracting the one or more features from the one or more dialog statements of the second span, includes classifying the one or more features to correspond with the at least one detector of the learning model.

4. The method of claim 1 , wherein the threshold condition is determined during training of the learning model to determine the topic change within the first span of the communication.

5. The method of claim 1 , wherein the one or more indicators of the topic change within the first span includes one or more of: a particular key phrase, a pause of particular duration, a particular activity on a participant's communication device, and a particular duration of the first span.

6. The method of claim 1 , wherein the at least one detector of the learning model includes a keyword and key phrase based detector, and at least one of: a time based detector, a personalized history based detector, a participant device activity based detector, and a text tiling based detector.

7. A computer program product for determining a topic change of a communication, the computer program product comprising:
a computer readable storage medium having program instructions embodied therewith, wherein the computer readable storage medium is not a transitory signal per se, the program instructions executable by a processor to cause the processor to perform a method comprising:
determining, a communication containing a set of dialog statements, wherein a first span of the communication includes one or more dialog statements of the set of dialog statements;
determining if the one or more dialog statements of the first span include one or more indicators of a topic change, wherein the one or more indicators are identified by at least one detector of a learning model;
responsive to determining the first span includes the one or more indicators of the topic change, generating a score for the one or more indicators, based on the learning model;
responsive to the score for the one or more indicators triggering a threshold condition, determining, by the computer processor, a topic change within the first span,
wherein the threshold condition determined during training of the learning model, includes:
determining, by the computer processor, a weighted value for the at least one detector, based on heuristics;
receiving input of labelled communication dialog statements, wherein the labelled communication dialog statements include one or more topic change indicators that are known, the one or more topic change indicators corresponding to the at least one detector;
adjusting, by the computer processor, the weighted value of the at least one detector in response to a delta between an output of scores of the at least one detector of the learning model and scores of the one or more topic change indicators that are known; and
determining, by the computer processor, the threshold condition in response to achieving an acceptable minimum for the delta between the output of the scores which are determined by the at least one detector of the learning model and the scores of the one or more topic change indicators that are known.

8. The computer program product of claim 7 , further comprising:
receiving the communication including the first span;
generating a second span based on adjusting boundaries of the first span by performing at least one of, adding to the first span, one or more dialog statements of the set of dialog statements not included in the first span, and removing one or more dialog statements from the first span;
determining a score for the first span and a score for the second span, wherein the score for the first span and the score for the second span is based on a topic of the first span and a topic of the second span, respectively;
responsive to the score of the second span being more favorable than the score of the first span, extracting one or more features from the one or more dialog statements of the second span not included in the first span; and
training the learning model to determine a topic change, based, at least in part, on including the one or more features from the one or more dialog statements of the second span, in at least one detector of the learning model.

9. The computer program product of claim 8 , wherein extracting the one or more features from the dialog statements of the second span, includes classifying the one or more features from the one or more dialog statements of the second span to correspond with the at least one detector of the learning model.

10. The c

Assignees

• IBM

Inventors

• Bhagwat Anand B
• Bohra Amit P
• Pal Manojkumar
• Shrivastava Abhishek
• Sahu Swapnasarit

Classifications

• G06F40/279

  Recognition of textual entities · CPC title

• G06N20/00 Primary

  Machine learning · CPC title

• G06F40/35

  Discourse or dialogue representation · CPC title

• G06F40/284

  Lexical analysis, e.g. tokenisation or collocates · CPC title

• G06F40/30

  Semantic analysis · CPC title