Arbitrary-Granularity Document Categorization With Proto-Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing document categorization systems face challenges in managing dynamic taxonomies, require significant computational resources for retraining, and are inefficient due to the need to re-evaluate documents when categories change, making them impractical for real-world applications.
Innovation Solution
A method and electronic device that uses proto-models, a compressed version of semantic models, to assign semantic categories based on document embeddings, allowing for flexible granularity and seamless addition or removal of categories without requiring a fixed taxonomy, and enabling high-precision detection and automatic model refinement.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a fixed taxonomy of categories is defined upfront for multi-class supervised learning, then the categorization system can be built with structured categories, but the system becomes complex to manage and requires significant time and resources to create adequate training datasets for each category
Solution Approach 1:
The patent segments the categorization task by using proto-models that represent compressed versions of semantic models for different categories. Instead of requiring a complete fixed taxonomy upfront, the system divides the categorization space into manageable proto-models that can be independently managed and updated, reducing the complexity of taxonomy management while maintaining categorization accuracy
Solution Approach 2:
The patent introduces dynamic adaptability by allowing the taxonomy to evolve without requiring complete redefinition. When new categories are needed, the system can add new proto-models or update existing ones dynamically, rather than requiring a fixed taxonomy to be redefined from scratch, thus reducing management complexity while maintaining reliability
2Adaptability or versatility
If the entire model is retrained from scratch when new categories are desired, then the model can accommodate new categories accurately, but the process becomes computationally expensive
Solution Approach 1:
The patent extracts the essential functionality of full model retraining by using proto-models as compressed representations. When new categories are added, only the relevant proto-models need to be updated or created rather than retraining the entire model, thus achieving category flexibility while significantly reducing computational resource consumption
Solution Approach 2:
The patent uses proto-models as compressed copies of semantic models. These proto-models capture the essential characteristics of categories in a space-efficient manner, allowing the system to adapt to new categories by working with these compressed representations rather than requiring full model retraining, thereby reducing computational costs while maintaining adaptability
3Reliability
If all documents are re-evaluated when new categories are created and models are updated, then the system can maintain accuracy with new categories, but the system becomes expensive and inefficient
Solution Approach 1:
The patent applies local quality by updating only the specific proto-models relevant to new categories rather than re-evaluating all documents against the entire taxonomy. This localized approach maintains categorization accuracy for affected categories while avoiding the inefficiency of re-evaluating all documents, thus improving system productivity without sacrificing reliability
4Ease of manufacture
If traditional ML approaches are used for document categorization, then the system can be built with word frequency features, but the system fails to handle synonymy and polysemy effectively due to word granularity uncertainty
Solution Approach 1:
The patent replaces traditional word-frequency-based mechanical approaches with deep learning models that process semantic representations. This substitution enables the system to handle synonymy and polysemy effectively by understanding contextual meaning rather than relying on word frequency counts, thereby improving measurement precision while maintaining ease of implementation through standardized deep learning frameworks
Data Source
AI summary
Embodiments herein disclose a method for determining at least one semantic category of at least one document using an electronic device 100. The method includes receiving at least one document embedding indicating a semantic representation of at least one document. Further, the method includes determining a probable set of semantic categories of a plurality of semantic categories associated with the document embedding based on an execution of the at least one document embedding on a plurality of proto-models. Further, the method includes receiving the semantic model associated with each of the probable set of semantic categories. Further, the method includes executing the at least one document embedding on the received semantic model. Further, the method includes determining the at least one semantic category out of the probable set of semantic categories, of the at least one document embedding based on the at least one executed document.