Hybrid learning system for natural language understanding

Patent metadata

Field	Value
Publication number	US-10497366-B2
Application number	US-201916238331-A
Country	US
Kind code	B2
Filing date	Jan 2, 2019
Priority date	Mar 23, 2018
Publication date	Dec 3, 2019
Grant date	Dec 3, 2019

Abstract

An agent automation system includes a memory configured to store a natural language understanding (NLU) framework, and a processor configured to perform actions, including: generating a meaning representation from an annotated utterance tree of an utterance, wherein a structure of the meaning representation indicates a syntactic structure of the utterance and one or more subtree vectors of the meaning representation indicate a semantic meaning of one or more intent subtrees of the meaning representation; searching the meaning representation of the utterance against an understanding model to extract intents/entities of the utterance based on the one or more subtree vectors of the meaning representation, wherein the understanding model includes a plurality of meaning representations derived from the intent/entity model; and providing the intents/entities of the utterance to a reasoning agent/behavior engine (RA/BE) of the agent automation system that performs one or more actions in response to the intents/entities of the utterance.

First claim

What is claimed is:

1. An agent automation system, comprising:
a memory configured to store an intent/entity model and a natural language understanding (NLU) framework, wherein the intent/entity model associates defined intents with sample utterances, and wherein the sample utterances encode defined entities as parameters of the defined intents within the intent/entity model; and
a processor configured to execute instructions of the NLU framework to cause the agent automation system to perform actions comprising:
generating a meaning representation from an annotated utterance tree of an utterance by combining a plurality of vectors of the annotated utterance tree using stored coefficients to calculate a respective subtree vector for each intent subtree of the annotated utterance tree, wherein each of the plurality of vectors is associated with a particular node or subtree that depends from each intent subtree, wherein the stored coefficients provide focus, attention, and magnification (FAM) to the calculation of each respective subtree vector, and wherein the meaning representation has a tree structure that indicates a syntactic structure of the utterance and has one or more subtree vectors indicating a semantic meaning of one or more intent subtrees of the meaning representation;
searching the meaning representation of the utterance against an understanding model to extract intents and entities of the utterance based on the one or more subtree vectors of the meaning representation, wherein the understanding model comprises a plurality of meaning representations derived from the sample utterances of the intent/entity model; and
providing the intents and entities of the utterance to a behavior engine (BE) of the agent automation system, wherein the BE is configured to perform one or more actions in response to the intents and entities of the utterance.

2. The system of claim 1 , wherein the NLU framework includes a meaning extraction subsystem that is configured to generate the meaning representation, and a meaning search subsystem that is configured to extract the intents and entities of the utterance and provide the intents and entities of the utterance to the BE.

3. The system of claim 1 , wherein the stored coefficients that provide the FAM to the calculation of each respective subtree vector comprise:
a first coefficient associated with a first type of node or subtree of the annotated utterance tree; and
a second coefficient associated with a second type of node or subtree of the annotated utterance tree, and wherein the first coefficient and the second coefficient have different values.

4. The system of claim 1 , wherein the plurality of vectors includes a word vector of a particular node of the annotated utterance tree that depends from the intent subtree.

5. The system of claim 1 , wherein the plurality of vectors include a second subtree vector generated by combining a second plurality of vectors using the stored coefficients, wherein each of the second plurality of vectors is associated with a second particular node or subtree that depends from the second subtree.

6. The system of claim 1 , wherein the NLU framework includes a meaning extraction subsystem having a vocabulary subsystem, a structure subsystem, and a prosody subsystem that cooperate to generate the annotated utterance tree of the utterance.

7. The system of claim 6 , wherein the structure subsystem and the prosody subsystem include a combination of machine learning (ML)-based and rules-based plugins that cooperate to parse the utterance into the annotated utterance tree.

8. The system of claim 1 , wherein, to search the meaning representation of the utterance, the processor is configured to execute instructions of the NLU framework to cause the agent automation system to perform actions comprising:
for each intent subtree of the at least one meaning representation of the utterance:
comparing the intent subtree to the plurality of meaning representations of the understanding model based on a compilation model template to generate a corresponding plurality intent subtree similarity scores; and

extracting the intents and entities of the utterance by determining which of the corresponding plurality of intent subtree similarity scores is highest or is above a predefined threshold value.

9. The system of claim 8 , wherein, to compare the intent subtree to the plurality of meaning representations of the understanding model, the processor is configured to execute instructions of the NLU system to cause the agent automation framework to perform actions comprising:
identifying a plurality of class compatible subtrees of the intent subtree and the plurality of meaning representations of the understanding model based on class compatibility rules of the compilation model template;
determining a respective class similarity score for each of the plurality of class compatible subtrees based on vector distances between subtree vectors associated with the plurality of class compatible subtrees; and
combining the respective class similarity score for each of the plurality of class compatible subtrees as a weighted average based on class-level scoring coefficients of the compilation model template to generate the corresponding plurality of intent subtree similarity scores.

10. The system of claim 1 , wherein the processor is configured to execute instructions of the NLU system to cause the agent automation system to perform actions comprising:
generating the plurality of meaning representations of the understanding model from the sample utterances of the intent/entity model before generating the meaning representation from the annotated utterance tree of the utterance.

11. A method of operating an agent automation system, comprising:
generating at least one meaning representation from an annotated utterance tree of an utterance by combining a plurality of vectors of the annotated utterance tree using stored coefficients to calculate a respective subtree vector for each intent subtree of the annotated utterance tree, wherein each of the plurality of vectors is associated with a particular node or subtree that depends from each intent subtree, wherein the stored coefficients provide focus, attention, and magnification (FAM) to the calculation of each respective subtree vector, and wherein the at least one meaning representation comprises a tree structure representative of a grammatical structure of the utterance and a plurality of subtree vectors representative of semantic meanings of a plurality of intent subtrees of the at least one meaning representation;
searching the at least one meaning representation of the utterance against an understanding model to extract intents and entities of the utterance, wherein the understanding model comprises plurality of meaning representations derived from sample utterances of an intent/entity model, wherein the intent/entity model associates defined intents with the sample utterances, and wherein the sample utterances encode defined entities as parameters of the defined intents within the intent/entity model; and
providing the intents and entities of the utterance to a behavior engine (BE) of the agent automation system to perform at least one action in response to the intents and entities of the utterance.

12. The method of claim 11 , wherein generating the at least one meaning representation of the utterance comprises:
generating a compilation unit triple for each of the plurality of intent subtrees of an annotated utterance tree of the utterance, wherein the compilation unit triple comprises a subtree vector of the plurality of subtree vectors, a first reference to a root no

Assignees

• Servicenow Inc

Inventors

• Sapugay Edwin
• Madamala Anil Kumar
• Naboka Maxim
• Sunkara Srinivas Satyasai
• Santos Lewis Savio Landry
• Subbarao Murali B

Classifications

• G10L15/1807

  using prosody or stress · CPC title

• G10L15/16

  using artificial neural networks · CPC title

• G06N3/006

  based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO] · CPC title

• G06N5/025

  Extracting rules from data · CPC title

• G06N20/00

  Machine learning · CPC title