Multiclassification Approach for Natural Language Classifiers

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Classification algorithms, such as Naive Bayes and neural networks, experience significant performance degradation as the number of classes and labels increase, leading to unacceptable turnaround times in time-sensitive operations, and traditional methods require resource-intensive expert curation and large amounts of high-confidence training statements.

Innovation Solution

The method involves determining an optimal cluster size for training sets, generating clusters, identifying high-confidence classes, creating dynamic models, and performing classifications using these models, which reduces processing requirements and execution duration without compromising accuracy or requiring additional training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If the number of classes and labels in classification algorithms is increased, then the classification coverage and capability are improved, but the performance degradation and turnaround time worsen

Engineering Contradiction:

Improveclassification coverageVSAvoidturnaround time

Core Design Contradiction:

Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent segments the large classification problem into multiple smaller sub-problems by creating clusters of related classes. Instead of classifying across all classes simultaneously, the system divides the class space into manageable clusters, classifies instances within each cluster, and then aggregates results. This segmentation reduces the computational complexity and turnaround time while maintaining comprehensive classification coverage across all classes.

Inventive Principle:

Principle #1Segmentation

2Reliability

If traditional classification methods are used with large numbers of classes, then comprehensive classification is achieved, but resource-intensive expert curation and large amounts of training data are required

Engineering Contradiction:

Improveclassification accuracyVSAvoidtraining data requirements

Core Design Contradiction:

ReliabilityVSQuantity of substance

Solution Approach 1:

The patent applies segmentation by dividing the training process into cluster-level and instance-level components. Rather than requiring extensive training data for every possible class combination, the system trains on cluster characteristics and then adapts to specific instances. This reduces the quantity of training data needed while maintaining high classification accuracy through the structured approach to handling multiple classes.

Inventive Principle:

Principle #1Segmentation

3Adaptability or versatility

If the number of classes in classification algorithms is increased, then more comprehensive categorization is achieved, but performance degradation occurs

Engineering Contradiction:

Improvecategorization comprehensivenessVSAvoidclassification performance

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent segments the classification task into hierarchical levels where broad categorization is performed at the cluster level and fine-grained classification occurs within clusters. This segmentation allows the system to handle comprehensive multi-class categorization while maintaining performance by reducing the effective class count at any given processing stage.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimensional organization by grouping classes into clusters, adding a hierarchical dimension to the classification process. This dimensional change allows the system to manage large numbers of classes by organizing them in a structured hierarchy, thereby maintaining performance while achieving comprehensive categorization.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

4Use of energy by moving object

If systems with limited memory and processing capabilities are used, then resource efficiency is improved, but classification execution speed may be affected

Engineering Contradiction:

Improvecomputational resource efficiencyVSAvoidclassification execution speed

Core Design Contradiction:

Use of energy by moving objectVSSpeed

Solution Approach 1:

The patent segments the classification computation into smaller, manageable tasks that can be executed efficiently on resource-constrained systems. By processing instances within clusters rather than across all classes simultaneously, the system reduces memory requirements and computational load at any given time, thereby enabling faster execution on systems with limited capabilities.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS11341358B2Multiclassification approach for enhancing natural language classifiers

Publication Date: 2022.05.24 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

In an approach to creating models utilizing optimally clustered training sets, one or more computer processors determine an optimal cluster size. The one or more computer processors generate one or more clusters from one or more classes and respectively associated training statements that are contained in a training set, based on the determined optimal cluster size, wherein the one or more generated clusters, respectively, contain fewer classes than the training set. The one or more computer processors identify one or more isolated high confidence classes and associated training statements from one or more cluster classifications generated by a static model trained with the one or more generated clusters. The one or more computer processors create one or more dynamic models trained with the one or more identified isolated high confidence classes. The one or more computer processors perform one or more classifications utilizing the one or more created dynamic models.