Conversation Model Training With Weighted Exemplars and Response Subsets

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing exemplar-based generative models often ignore or over-rely on retrieved responses, leading to uninteresting or overfitted responses, and are biased in learning scenarios with detailed time-order descriptions.

Innovation Solution

A method of training a conversation model by identifying a first context, selecting a response subset based on a gold response, and applying relevance scores to weight exemplars, excluding overly similar responses, to combine retrieval and generation models effectively.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a retrieval-based conversation model is used to extract responses from a predefined response set, then response diversity and interest are improved, but response accuracy and context matching may deteriorate

Engineering Contradiction:

Improveresponse diversityVSAvoidcontext matching accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent combines retrieval-based and generation-based conversation models into a unified system. The retrieval model extracts candidate responses from a predefined set, while the generation model refines these responses to ensure context accuracy. This merging allows the system to leverage both the diversity strength of retrieval models and the accuracy strength of generation models.

Inventive Principle:

Principle #5Merging (Combining)

2Ease of operation

If a generation-based conversation model is used to generate responses based on sequence to sequence architecture, then response fluency and naturalness are improved, but response diversity and interest deteriorate

Engineering Contradiction:

Improveresponse fluencyVSAvoidresponse diversity

Core Design Contradiction:

Ease of operationVSAdaptability or versatility

Solution Approach 1:

The system merges retrieval-based and generation-based models where the retrieval model provides diverse candidate responses and the generation model ensures fluent and natural output. This combination allows the system to achieve both diversity from the retrieval component and fluency from the generation component.

Inventive Principle:

Principle #5Merging (Combining)

3Adaptability or versatility

If existing exemplar-based generation models are used to provide retrieved responses as exemplars, then response diversity is improved, but learning bias towards detailed time-order descriptions occurs

Engineering Contradiction:

Improveresponse diversityVSAvoidlearning bias

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent modifies the training parameters and data representation to reduce bias towards detailed time-order descriptions. By changing how exemplars are selected and weighted during training, the system maintains response diversity while reducing learning bias towards specific description patterns.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS12475881B2Method of generating conversation information using examplar-based generation model and apparatus for the same

Publication Date: 2025.11.18 HYPERCONNECT INC

AI summary

A training method of a conversation model according to various example embodiments of the present disclosure may include identifying a first context, identifying a first response set corresponding to the first context based on a first model, identifying a response subset selected from the first response set based on a gold response corresponding to the first context and training a second model based on the first context information and the response subset.