Determining one or more topics of a conversation using a domain specific model

• US 8,626,509 B2
• Filed: 03/26/2008
• Issued: 01/07/2014
• Est. Priority Date: 10/13/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Roy, Shourya, Subramaniam, Laxminarayan Venkata
• Primary Examiner(s): Lerner, Martin
• Application Number:
  US12/055,653
• Publication Number:
  US 20080177538A1
• Time in Patent Office: 2,113 Days
• Field of Search: 704/235, 704/257, 704/275, 704/255, 706/55, 707/771, 707/777, 707/778
• US Class Current: 704/255
• CPC Class Codes: G10L 15/26 Speech to text systems G10L15/08 takes precedence

Abstract

Applications of a domain specific model are described. A domain specific model may encode information about a domain. Information available in the domain specific model may be used to identify a topic of a conversation, such as a topic of a call to a call center. Callers'"'"' complaints can be categorized into coarse as well as fine topic categories by analyzing an initial part of a call and by examining a distribution of topic specific descriptive and discriminative features within the initial portion of the call. Once a call has been identified as belonging to a topic, a call-center agent may be prompted with information about the topic, such as questions and answers and actions related to the topic. Generic to specific information may be provided to the agent as the call progresses.

Claims

1. A method comprising:

receiving a first set of one or more features of at least a first part of a conversation between at least a first speaker and a second speaker;

comparing, using at least one processor, the first set of one or more features of the at least the first part of the conversation to information regarding a node of a taxonomy of a domain specific model to determine whether the at least the first part of the conversation relates to a topic to which the node of the taxonomy corresponds, wherein the taxonomy of the domain specific model is hierarchical and comprises a plurality of nodes arranged in different levels of a hierarchy, wherein a higher-level node of the taxonomy corresponds to a broad topic and a lower-level node below the higher-level node in the taxonomy corresponds to a specific topic that relates to the broad topic and is more specific than the broad topic, and wherein comparing the first set of one or more features of the at least the first part of the conversation to the information regarding the node comprises comparing the first set of one or more features to information regarding the higher-level node of the taxonomy to determine whether the at least the first part of the conversation relates to the broad topic, wherein the comparing of the first set of one or more features to the information regarding the higher-level node has a first probability of providing a first prediction accuracy in identifying whether the at least the first part of the conversation relates to the broad topic;

in response to determining that a result of the comparing indicates that the at least the first part of the conversation relates to the broad topic, presenting, to a user interface of a device operated by the first speaker, topic-specific information related to the broad topic;

receiving a second set of one or more features of at least a second part of the conversation, the second part of the conversation including both the first part of the conversation and a subsequent part occurring in the conversation following the first part;

in response to determining that the result of the comparing of the first set of the one or more features to the information regarding the higher-level node indicates that the at least the first part of the conversation relates to the broad topic, comparing the second set of one or more features of the first part and the subsequent part to information regarding the lower-level node to determine whether the at least the second part of the conversation relate to the specific topic related to the broad topic, wherein the comparing of the second set of one or more features to the information regarding the lower-level node has a second probability of providing a second prediction accuracy in identifying whether the at least the second part of the conversation relates to the specific topic, the second prediction accuracy being higher than the first prediction accuracy; and

in response to determining that a result of the comparing of the second set of one or more features to the information regarding the lower-level node indicates that the at least the second part of the conversation relates to the specific topic, presenting to the first speaker topic-specific information related to the specific topic.

2. The method of claim 1, wherein receiving the one or more features of the at least the part of the conversation between at least the first speaker and the second speaker comprises receiving one or more features of at least a part of a conversation between a call-center agent and a caller regarding an issue, and

wherein presenting to the first speaker topic-specific information comprises, in response to determining that the result of the comparing indicates that the issue relates to the topic to which the node corresponds, presenting the topic-specific information to the call-center agent to aid the call-center agent in resolving the issue.

3. The method of claim 1, wherein presenting the topic-specific information to the first speaker comprises presenting suggestions to the first speaker of information to provide to the second speaker in the conversation and/or information to obtain from the second speaker in the conversation.

4. The method of claim 1, wherein:

comparing the first set of one or more features to the information regarding the higher-level node, presenting the topic-specific information related to the broad topic, comparing the second set of one or more features to the information regarding the lower-level node, and presenting topic-specific information related to the specific topic are carried out during the conversation,

comparing the first set of one or more features to information regarding the higher-level node and presenting the topic-specific information related to the broad topic are carried out after receipt of the first set of one or more features of the at least the first part of the conversation and before receipt of the second set of one or more features of the at least the second part of the conversation, and

comparing the second set of one or more features to the information regarding the lower-level node and presenting the topic-specific information related to the specific topic are carried out after receipt of the first set of one or more features of the at least the first part of the conversation and after receipt of the second set of one or more features of the at least the second part of the conversation.

5. The method of claim 1, wherein the node of the taxonomy is associated with at least one commonly-appearing feature commonly appearing in conversations related to the topic to which the node corresponds, the at least one commonly-appearing feature comprising one or more words and/or phrases, and

wherein comparing the first set of one or more features to the information regarding the node comprises comparing the first set of one or more features of the at least the part of the conversation to the at least one commonly-appearing feature associated with the node.

6. The method of claim 5, wherein:

the node of the taxonomy is further associated with at least one second feature that is more likely to appear in conversations related to the topic to which the node corresponds than to appear in conversations unrelated to the topic to which the node corresponds; and

wherein comparing the one or more features to the information regarding the node comprises comparing the one or more features of the at least the part of the conversation to the at least one commonly-appearing feature and to the at least one second feature associated with the node.

7. The method of claim 1, wherein the receiving, comparing, and presenting are performed on the device operated by the first speaker.

8. The method of claim 1, wherein:

the broad topic to which the higher-level node of the taxonomy corresponds is a first broad topic;

the specific topic to which the lower-level node corresponds is a first specific topic;

the taxonomy of the domain-specific model comprises at least one sibling node to the higher-level node at a same level in the hierarchy as the higher-level node, each of the at least one sibling node corresponding to at least one second broad topic;

the taxonomy of the domain-specific model comprises a plurality of child nodes of the higher-level node, the plurality of child nodes comprising the lower-level node and at least one second child node, each of the at least one second child node corresponding to at least one second specific topic related to the first broad topic to which the higher-level node corresponds;

the method further comprises determining a broad topic to which the at least the first part of the conversation relates, the determining comprising the comparing the first set of the one or more features of the at least the first part of the conversation to information regarding the higher-level node and comparing the first set of the one or more features to information regarding the at least one sibling node; and

the method further comprises determining a specific topic to which the at least the second part of the conversation relate, the determining comprising the comparing, in response to determining that the result of the comparing of the first set of the one or more features to the information regarding the higher-level node indicates that the at least the first part of the conversation relates to the first broad topic, the second set of the one or more features to the information regarding the lower-level node and comparing the second set of the one or more features to information regarding each of the at least one second child node of the higher-level node.

9. The method of claim 1, wherein:

the higher-level node of the taxonomy is associated with information comprising a first distribution of features in conversations related to the broad topic;

the lower-level node of the taxonomy is associated with information comprising a second distribution of features in conversations related to the specific topic;

comparing the first set of one or more features of the at least the first part of the conversation to the information regarding the higher-level node comprises comparing information regarding the first set of one or more features in the at least the first part of the conversation to the first distribution of features in conversations related to the broad topic to determine whether the at least the first part of the conversation relates to the broad topic; and

comparing the second set of one or more features of the at least the second part of the conversation to the information regarding the lower-level node comprises comparing information regarding the second set of one or more features of the at least the second part of the conversation to the second distribution of features in conversations related to the specific topic to determine whether the at least the second part of the conversation relate to the specific topic.

10. An apparatus comprising:

at least one processor; and

at least one computer-readable storage medium having encoded thereon processor-executable instructions that, when executed by the at least one processor, cause the at least one processor to carry out a method comprising;

receiving a first set of one or more features of at least a first part of a conversation between at least a first speaker and a second speaker;

comparing, using at least one processor, the first set of one or more features of the at least the first part of the conversation to information regarding a node of a taxonomy of a domain specific model to determine whether the at least the first part of the conversation relates to a topic to which the node of the taxonomy corresponds, wherein the taxonomy of the domain specific model is hierarchical and comprises a plurality of nodes arranged in different levels of a hierarchy, wherein a higher-level node of the taxonomy corresponds to a broad topic and a lower-level node below the higher-level node in the taxonomy corresponds to a specific topic that relates to the broad topic and is more specific than the broad topic, and wherein comparing the first set of one or more features of the at least the first part of the conversation to the information regarding the node comprises comparing the first set of one or more features to information regarding the higher-level node of the taxonomy to determine whether the at least the first part of the conversation relates to the broad topic, wherein the comparing of the first set of one or more features to the information regarding the higher-level node has a first probability of providing a first prediction accuracy in identifying whether the at least the first part of the conversation relates to the broad topic;

in response to determining that a result of the comparing indicates that the at least the first part of the conversation relates to the broad topic, presenting, to a user interface of a device operated by the first speaker, topic-specific information related to the broad topic;

receiving a second set of one or more features of at least a second part of the conversation, the second part of the conversation including both the first part of the conversation and a subsequent part occurring in the conversation following the first part;

in response to determining that the result of the comparing of the first set of the one or more features to the information regarding the higher-level node indicates that the at least the first part of the conversation relates to the broad topic, comparing the second set of one or more features of the first part and the subsequent part to information regarding the lower-level node to determine whether the at least the second part of the conversation relate to the specific topic related to the broad topic, wherein the comparing of the second set of one or more features to the information regarding the lower-level node has a second probability of providing a second prediction accuracy in identifying whether the at least the second part of the conversation relates to the specific topic, the second prediction accuracy being higher than the first prediction accuracy; and

in response to determining that a result of the comparing of the second set of one or more features to the information regarding the lower-level node indicates that the at least the second part of the conversation relates to the specific topic, presenting to the first speaker topic-specific information related to the specific topic.

11. The apparatus of claim 10, wherein:

comparing the first set of one or more features to the information regarding the higher-level node, presenting the topic-specific information related to the broad topic, comparing the second set of one or more features to the information regarding the lower-level node, and presenting topic-specific information related to the specific topic are carried out during the conversation,

comparing the first set of one or more features to information regarding the higher-level node and presenting the topic-specific information related to the broad topic are carried out after receipt of the first set of one or more features of the at least the first part of the conversation and before receipt of the second set of one or more features of the at least the second part of the conversation, and

comparing the second set of one or more features to the information regarding the lower-level node and presenting the topic-specific information related to the specific topic are carried out after receipt of the first set of one or more features of the at least the first part of the conversation and after receipt of the second set of one or more features of the at least the second part of the conversation.

12. The apparatus of claim 10, wherein presenting the topic-specific information to the first speaker comprises presenting suggestions to the first speaker of information to provide to the second speaker in the conversation and/or information to obtain from the second speaker in the conversation.

13. The apparatus of claim 10, wherein the apparatus is the device operated by the first speaker.

14. The apparatus of claim 10, wherein:

the broad topic to which the higher-level node of the taxonomy corresponds is a first broad topic;

the specific topic to which the lower-level node corresponds is a first specific topic;

the taxonomy of the domain-specific model comprises at least one sibling node to the higher-level node at a same level in the hierarchy as the higher-level node, each of the at least one sibling node corresponding to at least one second broad topic;

the taxonomy of the domain-specific model comprises a plurality of child nodes of the higher-level node, the plurality of child nodes comprising the lower-level node and at least one second child node, each of the at least one second child node corresponding to at least one second specific topic related to the first broad topic to which the higher-level node corresponds;

the method further comprises determining a broad topic to which the at least the first part of the conversation relates, the determining comprising the comparing the first set of the one or more features of the at least the first part of the conversation to information regarding the higher-level node and comparing the first set of the one or more features to information regarding the at least one sibling node; and

the method further comprises determining a specific topic to which the at least the second part of the conversation relate, the determining comprising the comparing, in response to determining that the result of the comparing of the first set of the one or more features to the information regarding the higher-level node indicates that the at least the first part of the conversation relates to the first broad topic, the second set of the one or more features to the information regarding the lower-level node and comparing the second set of the one or more features to information regarding each of the at least one second child node of the higher-level node.

15. The apparatus of claim 10, wherein:

the higher-level node of the taxonomy is associated with information comprising a first distribution of features in conversations related to the broad topic;

the lower-level node of the taxonomy is associated with information comprising a second distribution of features in conversations related to the specific topic;

comparing the first set of one or more features of the at least the first part of the conversation to the information regarding the higher-level node comprises comparing information regarding the first set of one or more features in the at least the first part of the conversation to the first distribution of features in conversations related to the broad topic to determine whether the at least the first part of the conversation relates to the broad topic; and

comparing the second set of one or more features of the at least the second part of the conversation to the information regarding the lower-level node comprises comparing information regarding the second set of one or more features of the at least the second part of the conversation to the second distribution of features in conversations related to the specific topic to determine whether the at least the second part of the conversation relate to the specific topic.

16. At least one computer-readable storage medium having encoded thereon processor-executable instructions that, when executed by at least one processor, cause the at least one processor to carry out a method comprising:

receiving a first set of one or more features of at least a first part of a conversation between at least a first speaker and a second speaker;

comparing, using at least one processor, the first set of one or more features of the at least the first part of the conversation to information regarding a node of a taxonomy of a domain specific model to determine whether the at least the first part of the conversation relates to a topic to which the node of the taxonomy corresponds, wherein the taxonomy of the domain specific model is hierarchical and comprises a plurality of nodes arranged in different levels of a hierarchy, wherein a higher-level node of the taxonomy corresponds to a broad topic and a lower-level node below the higher-level node in the taxonomy corresponds to a specific topic that relates to the broad topic and is more specific than the broad topic, and wherein comparing the first set of one or more features of the at least the first part of the conversation to the information regarding the node comprises comparing the first set of one or more features to information regarding the higher-level node of the taxonomy to determine whether the at least the first part of the conversation relates to the broad topic, wherein the comparing of the first set of one or more features to the information regarding the higher-level node has a first probability of providing a first prediction accuracy in identifying whether the at least the first part of the conversation relates to the broad topic;

in response to determining that a result of the comparing indicates that the at least the first part of the conversation relates to the broad topic, presenting, to a user interface of a device operated by the first speaker, topic-specific information related to the broad topic;

receiving a second set of one or more features of at least a second part of the conversation, the second part of the conversation including both the first part of the conversation and a subsequent part occurring in the conversation following the first part;

in response to determining that the result of the comparing of the first set of the one or more features to the information regarding the higher-level node indicates that the at least the first part of the conversation relates to the broad topic, comparing the second set of one or more features of the first part and the subsequent part to information regarding the lower-level node to determine whether the at least the second part of the conversation relate to the specific topic related to the broad topic, wherein the comparing of the second set of one or more features to the information regarding the lower-level node has a second probability of providing a second prediction accuracy in identifying whether the at least the second part of the conversation relates to the specific topic, the second prediction accuracy being higher than the first prediction accuracy; and

in response to determining that a result of the comparing of the second set of one or more features to the information regarding the lower-level node indicates that the at least the second part of the conversation relates to the specific topic, presenting to the first speaker topic-specific information related to the specific topic.

17. The at least one computer-readable storage medium of claim 16, wherein:

comparing the first set of one or more features to the information regarding the higher-level node, presenting the topic-specific information related to the broad topic, comparing the second set of one or more features to the information regarding the lower-level node, and presenting topic-specific information related to the specific topic are carried out during the conversation,

comparing the first set of one or more features to information regarding the higher-level node and presenting the topic-specific information related to the broad topic are carried out after receipt of the first set of one or more features of the at least the first part of the conversation and before receipt of the second set of one or more features of the at least the second part of the conversation, and

comparing the second set of one or more features to the information regarding the lower-level node and presenting the topic-specific information related to the specific topic are carried out after receipt of the first set of one or more features of the at least the first part of the conversation and after receipt of the second set of one or more features of the at least the second part of the conversation.

18. The at least one computer-readable storage medium of claim 16, wherein presenting the topic-specific information to the first speaker comprises presenting suggestions to the first speaker of information to provide to the second speaker in the conversation and/or information to obtain from the second speaker in the conversation.

19. The at least one computer-readable storage medium of claim 16, wherein presenting the topic-specific information to the first speaker comprises displaying the topic-specific information to the first speaker.

20. The at least one computer-readable storage medium of claim 16, wherein:

the broad topic to which the higher-level node of the taxonomy corresponds is a first broad topic;

the specific topic to which the lower-level node corresponds is a first specific topic;

the taxonomy of the domain-specific model comprises at least one sibling node to the higher-level node at a same level in the hierarchy as the higher-level node, each of the at least one sibling node corresponding to at least one second broad topic;

the taxonomy of the domain-specific model comprises a plurality of child nodes of the higher-level node, the plurality of child nodes comprising the lower-level node and at least one second child node, each of the at least one second child node corresponding to at least one second specific topic related to the first broad topic to which the higher-level node corresponds;

the method further comprises determining a broad topic to which the at least the first part of the conversation relates, the determining comprising the comparing the first set of the one or more features of the at least the first part of the conversation to information regarding the higher-level node and comparing the first set of the one or more features to information regarding the at least one sibling node; and

the method further comprises determining a specific topic to which the at least the second part of the conversation relate, the determining comprising the comparing, in response to determining that the result of the comparing of the first set of the one or more features to the information regarding the higher-level node indicates that the at least the first part of the conversation relates to the first broad topic, the second set of the one or more features to the information regarding the lower-level node and comparing the second set of the one or more features to information regarding each of the at least one second child node of the higher-level node.