Bridging Classifier for Web Query Taxonomy Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing query classification methods require extensive human-labeled training data and are inflexible when taxonomy changes occur, making them inefficient and costly, especially when using large intermediate taxonomies like the Open Directory Project.
Innovation Solution
A bridging classifier is built offline using an intermediate taxonomy to map queries to target categories, allowing for efficient online query processing and retraining only when necessary, with a category selection mechanism to reduce complexity by choosing relevant categories based on probability or mutual information scores.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If human-labeled training data is used for query classification, then classification accuracy is improved, but data collection time and cost increase significantly
Solution Approach 1:
The patent introduces an intermediate taxonomy as a mediator between the query and the target taxonomy. Instead of directly training a classifier on target taxonomy labels requiring human annotation, the system uses the intermediate taxonomy (with its own classifiers) to first map queries to intermediate categories, then transfers these mappings to the target taxonomy. This intermediary layer enables accurate classification without requiring extensive human-labeled training data for the target taxonomy itself.
Solution Approach 2:
The patent creates a copy of the classification task through the intermediate taxonomy. The intermediate taxonomy serves as a proxy or copy that can be trained independently using web crawl data and automatic classifiers. Once trained, this copy is used to infer classifications for the target taxonomy, effectively copying the classification function without requiring direct human labeling for the target categories.
2Adaptability or versatility
If the ODP taxonomy is used as intermediate taxonomy, then query mapping capability is improved, but computational complexity and cost increase due to 590,000+ categories
Solution Approach 1:
The patent segments the large ODP taxonomy into manageable components by selecting only the top N categories at each hierarchy level for use in the bridging classifier. Instead of processing all 590,000+ ODP categories, the system segments the taxonomy tree and uses only the most relevant segments (top categories) that provide sufficient mapping capability, thereby reducing computational complexity while maintaining query mapping effectiveness.
Solution Approach 2:
The patent applies partial action by using only a subset of the intermediate taxonomy categories rather than the complete taxonomy. The system selects top N categories at each level, which is a partial representation of the full ODP taxonomy. This partial approach provides sufficient query mapping capability without the excessive computational burden of processing all categories, achieving the right balance between capability and complexity.
3Measurement precision
If retraining is performed whenever target category structure changes, then classification accuracy is maintained, but system flexibility and operational stability deteriorate
Solution Approach 1:
The patent performs preliminary action by pre-computing and storing the relationship values between intermediate taxonomy categories and target taxonomy categories before online query processing. The bridging classifier is built offline using training data and saved relationship values, allowing the system to handle taxonomy changes without requiring retraining. When target categories change, the pre-built bridging classifier can adapt using the stored relationships rather than requiring complete retraining, thus maintaining accuracy while improving flexibility.
4Quantity of substance
If all intermediate taxonomy categories are used in the bridging classifier, then classification coverage is improved, but processing efficiency and relevance filtering deteriorate
Solution Approach 1:
The patent extracts only the relevant portion of the intermediate taxonomy by selecting top N categories at each hierarchy level. Instead of using all intermediate categories, the system extracts and uses only the most relevant ones that have the highest relationship values with target categories. This extraction improves processing efficiency by reducing the number of categories to process while maintaining adequate classification coverage through the selected top categories.
Data Source
AI summary
Described is a technology by which an intermediate taxonomy is processed (e.g., offline) with respect to a target taxonomy to determine relationship values between categories represented in the intermediate taxonomy and the target taxonomy. The relationship values are used to construct a bridging classifier for use in online query processing to relate queries to categories in the target taxonomy. The relation is based on each target category's relationship to one or more categories that were represented in the intermediate taxonomy. Further, only a relevant subset of the categories represented in the intermediate taxonomy may be chosen for use in the bridging classifier, e.g., based on relative probability scores and/or mutual information scores computed between the categories represented in the intermediate taxonomy and categories in the target taxonomy.