System and method for natural language understanding

Patent metadata

Field	Value
Publication number	US-11790895-B2
Application number	US-201916661581-A
Country	US
Kind code	B2
Filing date	Oct 23, 2019
Priority date	Jun 26, 2019
Publication date	Oct 17, 2023
Grant date	Oct 17, 2023

Abstract

An electronic device for natural language understanding includes at least one memory and at least one processor coupled to the at least one memory. The at least one processor is configured to process an utterance using a trained model. The at least one processor is also configured to replace a first portion of the utterance with a first token, where the first token represents a semantic role of the first portion of the utterance based on a slot vocabulary. The at least one processor is further configured to determine a slot value in the utterance based on the first token. In addition, the at least one processor is configured to perform a task corresponding to the utterance based on the determined slot value.

First claim

What is claimed is:

1. A method, comprising:
processing an utterance using a trained machine learning model;
partially delexicalizing the utterance using the trained machine learning model by replacing a first portion of the utterance with a first token from a slot vocabulary, wherein the first token represents a semantic role of the first portion of the utterance, and wherein the slot vocabulary includes a plurality of words or phrases each associated with one or more tokens included in the slot vocabulary;
determining, using the trained machine learning model, a slot value in the processed utterance based on the first token in the partially delexicalized utterance;
delexicalizing, in at least one candidate of one or more candidates that include the processed utterance, at least one word tagged with a slot type corresponding to a high variability slot type;
removing the at least one word from the at least one candidate based on a determination that a slot entropy score for the at least one candidate is above a threshold to create an altered candidate set;
determining, based on a calculated inverse of an average of one or more slot entropy scores, a parsing confidence score for each of the one or more candidates including the processed utterance;
selecting the processed utterance from among the one or more candidates based on the parsing confidence score for the processed utterance; and
performing a task corresponding to the utterance based on the determined slot value and the selected processed utterance.

2. The method of claim 1 , wherein:
the trained machine learning model is trained using training data including tokens and non-tokens for learning one or more semantic relationships between the tokens and the non-tokens; and
the slot vocabulary includes data different from the training data and is domain-dependent.

3. The method of claim 1 , further comprising:
replacing at least a second portion of the utterance with a second token prior to processing the utterance using the trained machine learning model.

4. The method of claim 3 , further comprising:
determining the second token based on string-matching of the second portion of the utterance and a content in the slot vocabulary.

5. The method of claim 1 , wherein replacing the first portion of the utterance with the first token includes:
identifying that a slot associated with the first portion of the utterance matches a predefined slot type; and
replacing, based on the identification that the slot matches the predefined slot type, the first portion of the utterance with the first token, wherein the first token corresponds to the predefined slot type.

6. The method of claim 1 , further comprising:
determining the utterance includes the first token; and
modifying the utterance such that a second portion of the utterance is replaced as part of the first token based on the one or more slot entropy scores.

7. An electronic device, comprising:
at least one memory; and
at least one processor coupled to the at least one memory, wherein the at least one processor is configured to:
process an utterance using a trained machine learning model;
replace, using the trained machine learning model, a first portion of the utterance with a first token from a slot vocabulary to partially delexicalize the utterance, wherein the first token represents a semantic role of the first portion of the utterance, and wherein the slot vocabulary includes a plurality of words or phrases each associated with one or more tokens included in the slot vocabulary;
determine, using the trained machine learning model, a slot value in the processed utterance based on the first token in the partially delexicalized utterance;
delexicalize, in at least one candidate of one or more candidates that include the processed utterance, at least one word tagged with a slot type corresponding to a high variability slot type;
remove the at least one word from the at least one candidate based on a determination that a slot entropy score for the at least one candidate is above a threshold to create an altered candidate set;
determine, based on a calculated inverse of an average of one or more slot entropy scores, a parsing confidence score for each of the one or more candidates including the processed utterance;
select the processed utterance from among the one or more candidates based on the parsing confidence score for the processed utterance; and
perform a task corresponding to the utterance based on the determined slot value and the selected processed utterance.

8. The electronic device of claim 7 , wherein:
the trained machine learning model is trained using training data including tokens and non-tokens for learning one or more semantic relationships between the tokens and the non-tokens; and
the slot vocabulary includes data different from the training data and is domain-dependent.

9. The electronic device of claim 7 , wherein the at least one processor is further configured to replace at least a second portion of the utterance with a second token prior to processing the utterance using the trained machine learning model.

10. The electronic device of claim 9 , wherein the at least one processor is further configured to determine the second token based on string-matching of the second portion of the utterance and a content in the slot vocabulary.

11. The electronic device of claim 7 , wherein, to replace the first portion of the utterance with the first token, the at least one processor is configured to:
identify that a slot associated with the first portion of the utterance matches a predefined slot type; and
replace, based on the identification that the slot matches the predefined slot type, the first portion of the utterance with the first token, wherein the first token corresponds to the predefined slot type.

12. The electronic device of claim 7 , wherein the at least one processor is configured to:
determine the utterance includes the first token; and
modify the utterance such that a second portion of the utterance is replaced as part of the first token based on the one or more slot entropy scores.

13. A non-transitory computer readable medium embodying a computer program, the computer program comprising instructions that when executed cause at least one processor of an electronic device to:
process an utterance using a trained machine learning model;
replace, using the trained machine learning model, a first portion of the utterance with a first token from a slot vocabulary to partially delexicalize the utterance, wherein the first token represents a semantic role of the first portion of the utterance, and wherein the slot vocabulary includes a plurality of words or phrases each associated with one or more tokens included in the slot vocabulary;
determine, using the trained machine learning model, a slot value in the processed utterance based on the first token in the partially delexicalized utterance;
delexicalize, in at least one candidate of one or more candidates that include the processed utterance, at least one word tagged with a slot type corresponding to a high variability slot type;
remove the at least one word from the at least one candidate based on a determination that a slot entropy score for the at least one candidate is above a threshold to create an altered candidate set;
determine, based on a calculated inverse of an average of one or more slot entropy scores, a parsing confidence score for each of the one or more candidates including the processed utterance;
select the processed utterance from among the one or more candidates based on the parsing confidence score for the pr

Assignees

• Samsung Electronics Co Ltd

Inventors

• Ray Avik
• Jin Hongxia

Classifications

• G10L15/1815 Primary

  Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning · CPC title

• G06F40/205

  Parsing · CPC title

• G06F40/284

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

• G06N20/00

  Machine learning · CPC title

• G10L15/1822 Primary

  Parsing for meaning understanding · CPC title