Synthetic Search Argument Generation for Sparse Product Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Search engines using artificial intelligence struggle to effectively train on infrequently searched products, leading to uneven training and poor user experience, as they require sufficient user interactions to learn correct interpretations of search arguments, which may not occur frequently enough for less popular items.
Innovation Solution
An apparatus and method generate synthetic search arguments to train the search engine, simulating various search terms and scenarios, including misspellings, abbreviations, and variations, to improve the search engine's ability to recognize and respond to diverse search queries without relying solely on user feedback.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the search engine relies on user interactions for training, then the training data reflects real user behavior, but infrequently searched products do not receive sufficient training data
Solution Approach 1:
The system performs preliminary action by generating synthetic search arguments in advance for infrequently searched products before actual user searches occur. This allows the search engine to be pre-trained on these products, ensuring sufficient training data is available even though real user interactions are sparse. The synthetic data generation proactively addresses the training data deficiency rather than waiting for natural user behavior to provide it.
Solution Approach 2:
Synthetic search arguments serve as an intermediary between the training system and the search engine. Instead of relying directly on sparse real user interactions, the system creates artificial intermediate training data that bridges the gap. These synthetic arguments simulate real user search behavior while ensuring adequate coverage for infrequently searched products, allowing the search engine to learn from this intermediate representation.
2Reliability
If the search engine waits for sufficient user interactions to train on a product, then training quality improves, but user experience deteriorates during the waiting period
Solution Approach 1:
The system applies preliminary action by generating synthetic search arguments and training the search engine on infrequently searched products in advance, before actual user searches occur. This eliminates the training time delay that would otherwise occur during user interactions, allowing the search engine to provide accurate responses immediately when users search for these products.
3Productivity
If the search engine uses synthetic search arguments for training, then training efficiency improves, but the complexity of the training system increases
Solution Approach 1:
A synthetic search argument generation system serves as an intermediary component that bridges the training data creation process and the search engine training process. This intermediary automatically generates realistic synthetic search arguments using product information and search patterns, making the increased system complexity worthwhile by enabling efficient training on infrequently searched products without manual intervention.
Data Source
AI summary
An apparatus and method for training a search engine generates synthetic search arguments and target information that are provided to the search engine as training. A search argument simulator ingests product information and historical interaction data, determines relevant descriptors, creates compound descriptors, and creates candidate search arguments. A trainer checks the candidates against historical search arguments, identifies similar search arguments and determines likely search arguments that are used for training. An evaluator monitors search engine responses and triggers retraining of the search engine.