Multimodal Content Classification With Hierarchical Zero-Shot Queries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional content classification models require full retraining when policies change, leading to unnecessary processing and lack of granularity, and they often produce binary outputs without detailed classification information.
Innovation Solution
Training individual large language models (LLMs) with domain-specific queries and hierarchical query structures to enable granular classification, allowing for easy policy updates and reduced processor utilization through zero-shot learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional binary content classifiers are used, then the system can filter certain categories from content, but the classification lacks granularity and requires full retraining when policies change
Solution Approach 1:
The patent segments the classification task into multiple hierarchical levels (e.g., general category, sub-category, specific topic). Instead of a single binary classifier, the system uses a hierarchy of classifiers where each level handles a specific aspect of classification. This segmentation allows policy changes to be applied at specific levels without retraining the entire classification system, reducing retraining time while maintaining adaptability.
Solution Approach 2:
The patent introduces a hierarchical dimension to the classification system, transforming the flat binary classification into a multi-level structured classification. By adding this dimensional structure, the system can apply policies at different levels independently, enabling flexible policy updates without full retraining while providing granular classification capabilities.
2Loss of information
If conventional binary classifiers are used, then the system can process content efficiently, but the output lacks detailed classification information
Solution Approach 1:
The classification system is segmented into multiple specialized classifiers, each responsible for a specific level of detail. This segmentation preserves rich classification information by distributing different aspects of classification across multiple components, while each individual classifier remains relatively simple in structure.
Solution Approach 2:
The hierarchical classifier system serves multiple functions simultaneously: it provides high-level filtering, detailed classification, and policy-based routing. This multi-functionality allows the system to maintain comprehensive classification information without requiring each individual component to be overly complex.
3Reliability
If full model retraining is performed when policies change, then the classifier adapts to new policies, but unnecessary processing occurs and computational overhead increases
Solution Approach 1:
The model is segmented into independent trainable modules corresponding to different classification levels and policies. When a policy changes, only the relevant segment needs retraining rather than the entire model, ensuring policy compliance while minimizing computational overhead and energy consumption.
Solution Approach 2:
Instead of performing full model retraining (excessive action), the system applies partial retraining only to the affected classifier modules. This partial action is sufficient to maintain policy compliance while significantly reducing the computational resources and energy required compared to complete retraining.
Data Source
AI summary
Systems and techniques for are described herein. A content passage is received from a corpus of training data comprising labeled training data and unlabeled training data. A query is received from a query hierarchy for a classification domain. The content passage and the query are embedded to form a passage-query pair. A predicted result for the passage-query pair is generated based on a calculated probability of the predicted result being within an answer threshold. A passage-query-result triplet is generated that comprises the passage-query pair and the predicted result according to the query hierarchy for the classification domain. Vectors of the content classification large language model are updated using the passage-query-result triplet.