Dynamic Intent Classification via Environment Variable Culling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current static intent classifiers in Intelligent Virtual Assistants (IVAs) face challenges in accurately selecting user intents from a large set of pre-defined classes, leading to increased processing time and decreased accuracy, especially when user eligibility filters out many intents, which are often not represented or poorly represented in the training data.
Innovation Solution
A dynamic intent classification system that uses environment variables to dynamically control the set of possible intents through culling rules, eliminating ineligible intents without retraining or redeploying the language model, by associating environment variables with ineligible intents and applying these culling rules to the classifier.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a static intent classifier is used with a large pre-defined set of intents, then the system can handle diverse user inputs, but the processing time increases and accuracy decreases when many intents are filtered out by user eligibility
Solution Approach 1:
The system performs preliminary filtering of intents based on user eligibility attributes (such as subscription status, location, device type) before the main classification process. This pre-filtering removes ineligible intents from consideration, so the classifier only processes a reduced subset of relevant intents, thereby reducing processing time without sacrificing accuracy
Solution Approach 2:
The intent classification process is segmented into two stages: (1) a filtering stage that divides the full intent set into eligible and ineligible groups based on user attributes, and (2) a classification stage that operates only on the eligible subset. This segmentation allows the system to maintain comprehensive intent coverage while optimizing processing efficiency
2Adaptability or versatility
If a static intent classifier is used with a large pre-defined set of intents, then the system can handle diverse user inputs, but the classification accuracy decreases when specific user behaviors are not well-represented in training data
Solution Approach 1:
The system applies different quality standards to different parts of the intent space. By identifying user-specific eligible intents based on local attributes (subscription type, location, device), the system ensures high classification accuracy for relevant intents while excluding irrelevant ones, rather than applying uniform classification to all intents
Solution Approach 2:
By pre-filtering intents based on user eligibility before classification, the system ensures that the classifier only attempts to classify intents that are actually relevant to the user. This preliminary action prevents the classifier from making inaccurate predictions on intents that should not apply to the user's context
3Reliability
If business rules require specific user behaviors to be mapped to pre-defined intents for regulatory compliance, then compliance is achieved, but the system becomes less flexible in handling evolving business requirements
Solution Approach 1:
The system dynamically adjusts the set of eligible intents based on user attributes and business rules. Rather than having a fixed mapping between user behaviors and intents, the system can adaptively include or exclude intents based on current user context and evolving business requirements, maintaining compliance while gaining flexibility
Solution Approach 2:
The system changes parameters of the classification process (specifically, the set of eligible intents) based on user attributes and business rules. By modifying which intents are available for classification rather than changing the intents themselves, the system maintains regulatory compliance while adapting to different user contexts and business requirements
Data Source
AI summary
To prevent intent classifiers from potentially choosing intents that are ineligible for the current input due to policies, dynamic intent classification systems and methods are provided that dynamically control the possible set of intents using environment variables (also referred to as external variables). Associations between environment variables and ineligible intents, referred to as culling rules, are used.