Ranker Training Data Generation via URL Product Name Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Websites face challenges in generating sufficient training data for customized computer-implemented rankers, leading to sub-optimal search results due to the lack of expertise and resources in searching and ranking technologies, especially for new websites with limited user interactions.
Innovation Solution
The method involves generating product name/query pairs by extracting product names from URLs in search logs, encoding them using word embeddings, and computing distances to assign queries to catalog product names, enabling the training of a customized ranker and enriching the searchable index with metadata.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a general-purpose ranker is used initially, then the website can provide search results without extensive training data, but the search results are sub-optimal and not customized for the website's catalog
Solution Approach 1:
The system performs preliminary actions by automatically collecting clickstream data from user interactions with search results, extracting product information from product detail pages, and generating synthetic training data before the ranker needs to be customized. This preliminary data collection and processing enables future customization without requiring extensive manual preparation or initial training data
Solution Approach 2:
The system implements self-service by automatically generating its own training data through web crawling and clickstream analysis without requiring external input or manual curation. The ranker system autonomously collects relevant product information, processes it through NLP techniques, and creates training datasets that will be used to customize the ranker for the specific website catalog
2Adaptability or versatility
If training data is collected from user interactions over time, then a customized ranker can be trained, but sub-optimal search results are returned during the data collection period
Solution Approach 1:
The system performs preliminary data collection and processing actions in advance by continuously crawling product detail pages, extracting product attributes, and building a product database before customization is needed. This head start reduces the time required to deploy a customized ranker when sufficient training data is eventually collected from user interactions
Solution Approach 2:
The system introduces an intermediary product database and NLP processing layer that bridges the gap between general-purpose search results and customized ranking. This intermediary structure allows the system to immediately begin collecting and processing relevant product information, preparing training data in parallel with the data collection period, thereby reducing the overall time to deployment
3Productivity
If product names are extracted from URLs using word embeddings, then training data can be generated without manual input, but the extraction accuracy depends on URL structure consistency
Solution Approach 1:
The system segments the product name extraction process into multiple stages: initial extraction from URL paths using word embeddings, validation through NLP techniques, and refinement by comparing against product catalog data. This segmented approach maintains high productivity through automation while improving accuracy through multiple verification steps
Solution Approach 2:
The system introduces an intermediary NLP processing layer that mediates between the automated URL-based product name extraction and the final training data. This intermediary layer uses natural language processing techniques to validate, correct, and refine extracted product names, ensuring accuracy while maintaining the productivity benefits of automated extraction
Data Source
AI summary
Technologies relating to generation of training data for use in training a computer-implemented ranker are described. A product name extracted from a URL in a log entry of a search log of a search engine is mapped to a catalog product name from a catalog of a website, wherein the catalog product name identifies a product that is available for acquisition by way of a webpage of a website. A query assigned to the URL in the log entry is assigned to the catalog product name due to the product name being mapped to the catalog product name. The computer-implemented ranker is trained based upon the query being assigned to the catalog product name.