Dialogue Model Training With Reference-Free Reply Discriminators

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing dialogue response generation systems face challenges in producing meaningful and logical responses due to reliance on external knowledge and large-scale training, which is resource-intensive and not sustainable, and fail to account for the variety of valid responses for a given context.

Innovation Solution

A neural network-based dialogue model is trained using reference-free discriminators that evaluate candidate replies based on specificity, consistency, fluency, and relatedness, determining quality scores through a weighted sum of discriminator scores, and employing a self-critical training algorithm to enhance response quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If external knowledge and large-scale training are used to improve response quality, then the dialogue model generates more meaningful responses, but resource consumption and training time increase significantly

Engineering Contradiction:

Improveresponse qualityVSAvoidresource consumption

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent extracts and removes the dependency on external knowledge bases and large-scale training data. Instead, it uses a self-contained discriminator model that evaluates candidate responses based on intrinsic quality metrics (specificity, consistency, fluency, relatedness) derived solely from the conversation history, eliminating the need for resource-intensive external resources

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The dialogue model serves itself by using the discriminator to evaluate its own generated candidate responses. The discriminator provides quality feedback that guides the selection of the best response without requiring external knowledge retrieval or large-scale pre-trained models, enabling the system to be self-sufficient and resource-efficient

Inventive Principle:

Principle #25Self-service

2Reliability

If external knowledge retrieval is used to enrich response features, then response quality improves, but error propagation from erroneous knowledge retrieval occurs

Engineering Contradiction:

Improveresponse qualityVSAvoiderror propagation

Core Design Contradiction:

ReliabilityVSObject-generated harmful factors

Solution Approach 1:

The patent converts the potential harm of erroneous external knowledge into a benefit by replacing external knowledge retrieval with an internal discriminator that evaluates responses based on conversation context. The discriminator's multi-dimensional quality assessment (specificity, consistency, fluency, relatedness) transforms the problem of knowledge accuracy into a benefit of context-aware response evaluation that is immune to external knowledge errors

Inventive Principle:

Principle #22Blessing in disguise (Convert harm into benefit)

3Adaptability or versatility

If large-scale training is performed to handle language evolution, then model performance improves, but training time and computational cost increase

Engineering Contradiction:

Improvelanguage evolution handlingVSAvoidtraining time

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent introduces dynamic adaptability through the discriminator's real-time evaluation of candidate responses against conversation history. The system adapts to language evolution and new topics dynamically during inference by evaluating responses based on their fluency, consistency, and relatedness to the current conversation context, eliminating the need for static large-scale retraining

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent changes the evaluation parameters from external knowledge accuracy to intrinsic response quality metrics (specificity, consistency, fluency, relatedness). This parameter transformation allows the system to handle language evolution and new topics by evaluating responses based on their contextual appropriateness rather than relying on pre-trained knowledge, enabling rapid adaptation without computational retraining

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20260017516A1Dialogue model training based on reference-free discriminators

Publication Date: 2026.01.15 TENCENT AMERICA LLC

AI summary

A method of generating a neural network based open-domain dialogue model, includes receiving an input utterance from a device having a conversation with the dialogue model, obtaining a plurality of candidate replies to the input utterance from the dialogue model, determining a plurality of discriminator scores for the candidate replies based on reference-free discriminators, determining a plurality of quality score associated with the candidate replies, and training the dialogue model based on the quality scores.