Hierarchical Intent Clustering for NLU Precision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As the number of user intents in natural language understanding (NLU) systems grows, intent classification performance degrades due to overlapping intents, and existing hierarchical approaches require significant manual effort and cost to set up.
Innovation Solution
A hierarchical approach using a well-balanced supermodel and clusters of models trained with extra utterances, which automatically identifies intent clusters using an unsupervised machine-learning algorithm, reducing the need for manual ontology definition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a single model is trained to process all intents, then the system can handle a comprehensive set of user intents, but the intent classification performance degrades due to overlapping intents
Solution Approach 1:
The patent divides the single intent classification model into multiple specialized models, each trained on a specific subset of intents. This segmentation allows each model to focus on a narrower scope, reducing intent overlap and improving classification precision while collectively maintaining comprehensive intent coverage across all models.
2Measurement precision
If additional datasets are added to train the model with more specificity, then the model can distinguish between similar intents, but more overlapping intents are introduced
Solution Approach 1:
Instead of adding more data to a single model, the patent segments the training data into multiple subsets, each corresponding to a specific intent cluster. This allows specialized models to be trained on focused datasets, achieving high precision for specific intent distinctions without introducing harmful overlaps that would occur in a monolithic model.
3Measurement precision
If a hierarchical structure is manually set up with domains and ontologies, then intent classification performance is improved, but significant cost and manual effort are required
Solution Approach 1:
The patent employs unsupervised machine learning algorithms that automatically cluster intents and train specialized models without requiring manual ontology definition or hierarchical structure setup. The system self-organizes the intent classification architecture, eliminating the need for expensive manual configuration while maintaining improved classification performance through specialized models.
4Measurement precision
If the system is overtrained for nearly all exactly possible utterances, then intent classification accuracy is improved, but overall system performance degrades
Solution Approach 1:
The patent segments the overtraining problem into manageable portions by creating multiple specialized models, each trained on a specific subset of intents. This allows each model to achieve high accuracy for its specific domain without the computational burden and performance degradation associated with overtraining a single model on all possible utterances.
Data Source
AI summary
The present disclosure relates generally to natural language understanding. An exemplary computer-enabled method comprises: receiving first utterance data corresponding to a first plurality of intents; identifying a first plurality of intent clusters using an unsupervised machine-learning algorithm, wherein each intent cluster of the first plurality of intent clusters comprises a respective subset of the first plurality of intents; training, based on the first utterance data, an intent cluster classification model, wherein the intent cluster classification model is configured to receive a user utterance and identify an intent cluster of the first plurality of intent clusters; and training, based on the first utterance data, an intent classification model for each intent cluster of the first plurality of intent clusters to obtain a plurality of intent classification models, wherein each intent classification model is configured to receive the user utterance and identify an intent from the respective intent cluster.