Granted Patent B2

US 12,475,881 · App. 17/805,191 · Granted Nov 18, 2025

Inventors: Enkhbayar Erdenee (Seoul, KR); Beom Su Kim (Seoul, KR); Seok Jun Seo (Seoul, KR); Sang Il Ahn (Cheongju-si, KR); Bu Ru Chang (Seoul, KR); Seung Ju Han (Seoul, KR)

Assignee: Hyperconnect LLC

G10L15/063G10L15/16

Abstract

Claims (30)

1 . A method of training a conversation model in an electronic apparatus, the method comprising:

identifying a first context;

identifying a first response set corresponding to the first context based on a first model;

excluding, from the first response set, at least one response that exists outside a first range in an embedding space from a gold response corresponding to the first context, wherein the first range is determined based on a k-means algorithm or a k-Nearest Neighbor (kNN) algorithm;

excluding, from the first response set, at least another response that is included in a second range in the embedding space from the gold response, wherein the second range indicates a Jaccard Filter Boundary in the embedding space;

identifying a response subset including a plurality of responses selected from the first response set based on the gold response;

calculate relevance scores, each relevance score indicating a degree of relevance to the gold response for a response in the plurality of responses; and

training a second model based on the first context, the response subset, and the relevance scores.

2 . The method of claim 1 , wherein the second model identifies a second context for a conversation obtained from a user, and provides a gold response for the second context based on the second context.

3 . The method of claim 1 , wherein the first context includes one or more conversations obtained from a user.

4 . The method of claim 1 , wherein the second model is trained by performing a backpropagation operation using a loss function calculated based on a set of weights, wherein each weight of the set of weights is calculated by normalizing the relevance scores for each response.

5 . An electronic apparatus for performing a method of generating a conversation, the electronic apparatus comprising:

a memory in which at least one program is stored; and

a processor, wherein the processor is configured to identify a first context;

identify a first response set corresponding to the first context based on a first model;

identify a gold response corresponding to the first context based on a second model;

exclude, from the first response set, at least one response that exists outside a first range in an embedding space from the gold response, wherein the first range is determined based on a k-means algorithm or a k-Nearest Neighbor (kNN) algorithm;

exclude, from the first response set, at least another response that is included in a second range in the embedding space from the gold response, wherein the second range indicates a Jaccard Filter Boundary in the embedding space;

identify a response subset including a plurality of responses selected from the first response set based on the gold response;

calculate relevance scores, each relevance score indicating a degree of relevance to the gold response for a response in the plurality of responses; and

train the second model based on the first context, the response subset, and the relevance scores.

6 . A non-transitory computer-readable recording medium, comprising a medium configured to store computer readable instructions, wherein, when the computer readable instructions are executed by a processor, the processor performs operations comprising:

identifying a first context;

identifying a first response set corresponding to the first context based on a first model;

identifying a gold response corresponding to the first context based on a second model;

excluding, from the first response set, at least one response that exists outside a first range in an embedding space from the gold response, wherein the first range is determined based on a k-means algorithm or a k-Nearest Neighbor (kNN) algorithm;

excluding, from the first response set, at least another response that is included in a second range in the embedding space from the gold response, wherein the second range indicates a Jaccard Filter Boundary in the embedding space;

identifying a response subset including a plurality of responses selected from the first response set based on the gold response;

calculating relevance scores, each relevance score indicating a degree of relevance to the gold response for a response in the plurality of responses; and

training the second model based on the first context, the response subset, and the relevance scores.

Assignments (1)

Priority Claims (2)

Continuity (1)

References Cited (177)