Dynamic Conversational Response Training via Intent Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The collection of training data for sequential conversational responses is challenging due to the need for diverse and extensive user interactions to cover the universe of possible intents, which results in a large and complex data set requiring a long experimental process affecting a large user base.
Innovation Solution
The system employs a novel mechanism using multiple machine learning models to determine intent clusters, where a first model clusters specific intents into clusters through unsupervised hierarchical clustering, and a second model selects a subset of intent clusters for display, leveraging user actions to label and generate dynamic conversational responses across clusters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If continuous recommendations are generated to initiate user responses to observe both negative and positive outcomes, then adequate training data can be collected, but the experimental process becomes long and affects a large set of users
Solution Approach 1:
The patent applies preliminary action by using unsupervised hierarchical clustering to pre-group intents into clusters before the actual data collection process. This preprocessing step organizes the intent universe in advance, allowing the system to efficiently select representative intents from each cluster for experimentation, thereby reducing the overall experimental duration while maintaining training data quality.
Solution Approach 2:
The patent implements partial action by selecting only a subset of representative intents from each cluster for the experimental process, rather than testing all possible intents. This partial sampling approach reduces the number of users affected and shortens experimental duration while still gathering sufficient training data through the clustering-based generalization to other intents in the same cluster.
2Adaptability or versatility
If the size of the intents universe is large covering hundreds of intents, then comprehensive training coverage is achieved, but the number of possible sequences explodes to millions requiring extensive data collection
Solution Approach 1:
The patent applies segmentation by dividing the large intent universe into smaller, manageable clusters using unsupervised hierarchical clustering. This segmentation groups similar intents together, allowing the system to handle hundreds of intents by processing them in organized clusters rather than as individual items, thereby reducing the complexity of data collection while maintaining comprehensive intent coverage.
Solution Approach 2:
The patent implements universality by training the machine learning model on sequences from representative intents within each cluster, and then generalizing this training to cover all intents in the same cluster. This multi-functional approach allows a single training process to serve multiple intents simultaneously, reducing data collection complexity while maintaining comprehensive coverage across the entire intent universe.
3Measurement precision
If adequate learning sample sizes are obtained for the intents' universe, then machine learning prediction accuracy improves, but the experimental process requires a long duration affecting large user base
Solution Approach 1:
The patent applies copying by using the training results from representative intents within each cluster and generalizing them to cover all intents in the same cluster. Instead of collecting separate training data for every single intent, the system copies the learning outcomes from representative samples to their corresponding cluster members, thereby achieving adequate learning sample sizes across the entire intent universe with significantly reduced data collection effort and improved productivity.
Data Source
AI summary
Methods and systems for generating training data for sequential conversational responses to alleviate the collection burden are disclosed herein. More specifically, the methods and systems describe generating determining dynamic conversational responses to generate training data for sequential conversation responses. For example, the system creates a special form of a random treatment control experiment to reduce the data collection while at the same time generating effectively adequate learning data for intents for the sequenced machine learning.