Granted Patent B2

US 12,566,921 · App. 18/087,629 · Granted Mar 3, 2026

Inventors: Tuyen Quang Pham (Melbourne, AU); Cong Duy Vu Hoang (Melbourne, AU); Mark Edward Johnson (Sydney, AU); Thanh Long Duong (Melbourne, AU)

Assignee: Oracle International Corporation

G06F40/295G06F40/205G06F40/284

Abstract

Claims (71)

1 . A computer-implemented method comprising:

receiving an utterance from a user;

inputting, by a deep learning named entity recognition (NER) subsystem, the utterance into a neural network comprising a Bidirectional Long Short-Term Memory Conditional Random Field (BiLSTM-CRF) architecture and model parameters learned for named entity recognition;

generating, by the neural network of the NER subsystem, a first representation of one or more named entities based on the utterance;

searching, by an elastic search subsystem, a gazetteer, comprising executing an elastic entity search on the gazetteer based on the utterance to generate an initial search result;

multi-hot encoding, by the elastic search subsystem, the initial search result to generate a second representation of the one or more named entities identified in the utterance, wherein multi-hot encoding the initial search result includes generating the second representation that includes two or more nonzero terms based on a relative location of tokens in the initial search result;

combining, by a conditional random field (CRF) layer of the BiLSTM-CRF, the first representation with the second representation to generate a combined named entity representation;

outputting, by the CRF layer, the combined named entity representation; and

based on the combined named entity representation, generating a response and providing the response to the user as speech output via a speaker and/or text output via a display.

2 . The method of claim 1 , wherein combining the first representation with the second representation comprises:

computing a weighted linear interpolation of the first representation and the second representation.

3 . The method of claim 1 , wherein combining the first representation with the second representation comprises:

concatenating the first representation and the second representation.

4 . The method of claim 1 , wherein combining the first representation with the second representation comprises:

computing a sequence of attention vectors based on the first representation and the second representation.

5 . The method of claim 1 , wherein the method is performed by a machine learning model comprising the neural network, the method further comprising training the machine learning model by:

accessing a training utterance;

inputting the training utterance into the neural network;

generating, by the neural network, a first training representation based on the training utterance;

searching the gazetteer based on the training utterance to generate a second training representation; and

training the machine learning model based on the first training representation and the second training representation.

6 . The method of claim 5 , wherein training the machine learning model further comprises:

identifying a dropout parameter; and

applying the dropout parameter to the first training representation.

7 . A system comprising:

one or more processors; and

a memory coupled to the one or more processors, the memory comprising a plurality of instructions executable by the one or more processors to cause the one or more processors to perform operations comprising:

receiving an utterance from a user;

inputting, by a deep learning named entity recognition (NER) subsystem, the utterance into a neural network comprising a Bidirectional Long Short-Term Memory Conditional Random Field (BiLSTM-CRF) architecture and model parameters learned for named entity recognition;

generating, by the neural network, a first representation of one or more named entities based on the utterance;

searching, by an elastic search subsystem, a gazetteer, comprising executing an elastic entity search on the gazetteer based on the utterance to generate an initial search result;

multi-hot encoding, by the elastic search subsystem, the initial search result to generate a second representation of the one or more named entities identified in the utterance, wherein multi-hot encoding the initial search result includes generating the second representation that includes two or more nonzero terms based on a relative location of tokens in the initial search result;

combining, by a conditional random field (CRF) layer of the BiLSTM-CRF, the first representation with the second representation to generate a combined named entity representation;

outputting, by the CRF layer, the combined named entity representation; and

based on the combined named entity representation, generating a response and providing the response to the user as speech output via a speaker and/or text output via a display.

8 . The system of claim 7 , wherein the operation of combining the first representation with the second representation comprises:

computing a weighted linear interpolation of the first representation and the second representation.

9 . The system of claim 7 , wherein the operation of combining the first representation with the second representation comprises:

concatenating the first representation and the second representation.

10 . The system of claim 7 , wherein the operation of combining the first representation with the second representation comprises:

computing a sequence of attention vectors based on the first representation and the second representation.

11 . The system of claim 7 , wherein the operations are performable by a machine learning model comprising the neural network, and wherein the operations further comprise training the machine learning model by:

accessing a training utterance;

inputting the training utterance into the neural network;

generating, by the neural network, a first training representation based on the training utterance;

searching the gazetteer based on the training utterance to generate a second training representation; and

training the machine learning model based on the first training representation and the second training representation.

12 . The system of claim 11 , wherein the operation of training the machine learning model further comprises:

identifying a dropout parameter; and

applying the dropout parameter to the first training representation.

13 . A non-transitory computer-readable memory comprising a plurality of instructions executable by one or more processors to cause the one or more processors to perform operations comprising:

receiving an utterance from a user;

inputting, by a deep learning named entity recognition (NER) subsystem, the utterance into a neural network comprising a Bidirectional Long Short-Term Memory Conditional Random Field (BiLSTM-CRF) architecture and model parameters learned for named entity recognition;

generating, by the neural network of the NER subsystem, a first representation of one or more named entities based on the utterance;

searching, by an elastic search subsystem, a gazetteer, comprising executing an elastic entity search on the gazetteer based on the utterance to generate an initial search result;

multi-hot encoding, by the elastic search subsystem, the initial search result to generate a second representation of the one or more named entities identified in the utterance, wherein multi-hot encoding the initial search result includes generating the second representation that includes two or more nonzero terms based on a relative location of tokens in the initial search result;

combining, by a conditional random field (CRF) layer of the BiLSTM-CRF, the first representation with the second representation to generate a combined named entity representation;

outputting, by the CRF layer, the combined named entity representation; and

based on the combined named entity representation, generating a response and providing the response to the user as speech output via a speaker and/or text output via a display.

14 . The non-transitory computer-readable memory of claim 13 , wherein the operation of combining the first representation with the second representation comprises:

computing a weighted linear interpolation of the first representation and the second representation.

15 . The non-transitory computer-readable memory of claim 13 , wherein the operation of combining the first representation with the second representation comprises:

concatenating the first representation and the second representation.

16 . The non-transitory computer-readable memory of claim 13 , wherein the operation of combining the first representation with the second representation comprises:

computing a sequence of attention vectors based on the first representation and the second representation.

17 . The non-transitory computer-readable memory of claim 13 , wherein the operations are performable by a machine learning model comprising the neural network, and wherein the operations further comprise training the machine learning model by:

accessing a training utterance;

inputting the training utterance into the neural network;

generating, by the neural network, a first training representation based on the training utterance;

searching the gazetteer based on the training utterance to generate a second training representation; and

training the machine learning model based on the first training representation and the second training representation.

Assignments (1)

Continuity (2)

References Cited (26)