Iterative Intent Builder with Dynamic Batch Scheduling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current automated response systems, such as chatbots, face challenges in accurately interpreting user intents due to the complexity and variability of human queries, leading to suboptimal intent recognition performance, especially when trained with limited and labor-intensive annotated data.
Innovation Solution
The method involves utilizing batch utterance scheduling with two computational models to iteratively build intents, dynamically scheduling jobs based on statistical criteria, and seamlessly integrating results from short- and long-running jobs to provide continuous and relevant recommendations, allowing for real-time updates without requiring full model retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If full model retraining is performed to improve intent classification accuracy, then recognition performance improves, but processing time and computational resources increase significantly
Solution Approach 1:
The patent segments the intent building process into iterative batches, where only subsets of utterances are processed in each iteration rather than complete retraining. The intent builder divides training data into manageable batches that can be processed incrementally, allowing the system to update intent classifications without performing full model retraining on the entire dataset.
Solution Approach 2:
The system implements dynamic scheduling of batch processing jobs based on statistical criteria and system state. The intent builder dynamically adjusts processing parameters, batch sizes, and scheduling decisions to optimize between accuracy improvement and processing time, allowing the system to adapt its behavior based on current performance metrics and resource availability.
2Measurement precision
If more annotated training data is collected to improve intent recognition, then classification accuracy improves, but data preparation time and labor increase
Solution Approach 1:
The system performs preliminary processing of unlabeled utterances by separating them into subsets and pre-processing them according to different models before full intent building. This preliminary action prepares data in advance for more efficient processing during intent classification, reducing the need for manual annotation while maintaining accuracy.
Solution Approach 2:
The intent builder automatically retrieves and processes additional examples from data sources without requiring manual data collection or annotation. The system self-services by automatically gathering relevant utterances, separating them into appropriate subsets, and incorporating them into training iterations, eliminating the need for external data preparation labor.
3Measurement precision
If complex processing models are used to handle query variability, then intent interpretation accuracy improves, but computational complexity increases
Solution Approach 1:
The patent segments the processing of variable queries into different subsets handled by different models. Rather than using a single complex model for all queries, the system divides utterances into first and second subsets processed by different computational approaches, allowing each model to specialize in handling specific types of query variability.
Solution Approach 2:
The system applies complex processing selectively to specific subsets of data rather than applying maximum complexity to all data. By processing only the second subset according to the second model and incorporating results iteratively, the system achieves high accuracy for critical cases while avoiding unnecessary computational complexity for other cases.
Data Source
AI summary
Expansion of intent classification data utilizing batch utterance scheduling, by a processor in a computing environment. A set of unlabeled examples for intent processing is received by an intent builder iteratively defining an intent. The set of examples are separated into a first subset processed according to a first model and a second subset processed according to a second model. The first subset is incorporated into the intent builder during a building iteration and scheduling a first batch processing of the second subset processed according to the second model based on a scheduling criteria. The first batch processing of the second subset is initiated once the scheduling criteria is satisfied. Upon completion of the first batch processing, results of the completion are used to influence additional examples retrieved from the first subset and the second subset during a subsequent building iteration by the intent builder.