Intent Recognition Without Retraining

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Traditional machine learning-based dialog systems require frequent retraining when intents or values change, making it cumbersome and time-consuming to update and deploy voice-enabled systems with new functionalities.

Innovation Solution

The use of a match function that generates a vector representation of spoken utterances and calculates match scores for key-value pairs, allowing for immediate identification of intents or answers without the need for retraining when values are added or removed, using techniques such as cosine similarity, bilinear functions, or neural networks.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If traditional machine learning-based techniques are used for intent recognition, then the system can accurately predict intents from preconfigured sets, but frequent retraining is required when intents or values change, making the system cumbersome and time-consuming to update

Engineering Contradiction:
Improveability to update intents and valuesVSAvoidtime required for retraining
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent extracts the training process from the operational system by using a separate training dataset that is distinct from the production data. The training data includes explicit intent labels and value associations, allowing the system to learn mappings without requiring retraining of the entire model when new intents or values are added. This separation enables incremental updates to the knowledge base without triggering full retraining cycles.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system performs preliminary actions by pre-processing training data to create structured associations between intents, values, and example utterances before deployment. The training phase establishes foundational knowledge structures that allow the operational system to handle new intents and values through simple data additions rather than complex retraining procedures.

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If traditional machine learning-based techniques are used for intent recognition, then the system can handle complex natural language inputs, but a great deal of computing resources must be devoted to frequent retraining procedures

Engineering Contradiction:
Improveability to add new functionalitiesVSAvoidcomputational resources for retraining
Core Design Contradiction:
Adaptability or versatilityVSUse of energy by moving object

Solution Approach 1:

The patent extracts the computationally intensive training operations from the operational system by maintaining a separate training dataset with explicit labels. New functionalities are added by simply adding new labeled examples to the training data rather than retraining the entire model, dramatically reducing computational resource requirements for updates.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

Instead of performing full retraining when new intents or values are added, the system applies partial updates by adding only the necessary new training examples and performing targeted training on the specific new functionalities. This partial action approach consumes significantly fewer computational resources while achieving the same adaptability goals.

Inventive Principle:
Principle #16Partial or excessive action

3Measurement precision

If classifier-based approaches are used to predict answers for input questions, then the system can provide accurate responses, but the system must be retrained whenever underlying data or values of interest change

Engineering Contradiction:
Improveaccuracy of answer predictionVSAvoidease of updating values
Core Design Contradiction:
Measurement precisionVSEase of operation

Solution Approach 1:

The patent segments the knowledge base into distinct components: intents, values, and their associations. Each component can be independently updated in the training dataset without affecting others. This segmentation allows accurate answer prediction to be maintained while simplifying the update process, as new values can be added to the training data without requiring system-wide retraining.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system performs preliminary structuring of data relationships during the training phase, establishing clear mappings between intents, values, and example questions. This preliminary organization enables easy updates to underlying data, as new values can be integrated into the existing structured framework without complex retraining procedures.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4028932B1Reduced training intent recognition techniques
Publication Date: 2024.06.19 ORACLE INT CORP
  • EP4028932B1 patent drawingFigure 1
  • EP4028932B1 patent drawingFigure 2
  • EP4028932B1 patent drawingFigure 3

AI summary

The present disclosure relates generally to determining intent based upon speech input using a dialog system. More particularly, techniques are described using matching-based machine learning techniques to identify an intent corresponding to speech input in a dialog system. These procedures do not require training when intents are added or removed from the set of possible intents.