Transformer Search Ranking with Two-Phase Interaction Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current web search systems face challenges in effectively ranking search results using transformer-based models due to the inclusion of non-text data, such as user click information, which complicates training and reduces the efficiency of relevance determination.
Innovation Solution
A two-phase training approach for transformer-based learning models, where the first phase uses a broader data set to predict user interactions without indicating positive interactions, and the second phase refines the model using only positive interactions, allowing for improved personalized ranking.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If transformer-based models are trained using traditional single-phase methods with mixed data including non-text data, then the model can process diverse search data, but the training efficiency and relevance determination accuracy deteriorate
Solution Approach 1:
The training process is divided into two distinct phases: Phase 1 trains the transformer model using only text-based query-document pairs to establish fundamental language understanding, while Phase 2 introduces non-text data (click information, user interactions) as additional features for personalized ranking. This segmentation allows the model to first master core NLP capabilities before layering on complex interaction signals, thereby improving training efficiency while maintaining adaptability to diverse data types.
Solution Approach 2:
The model performs preliminary training on text-only data before incorporating non-text interaction data. This preliminary action establishes a solid foundation of language comprehension and semantic understanding, which then serves as the base for subsequent personalized ranking improvements. The preliminary text-based training ensures the model can process diverse search data effectively before the complexity of multi-phase interaction modeling is introduced.
2Quantity of substance
If all available training data including non-text click information is used in single-phase training, then data utilization is maximized, but the model's ability to capture semantic understanding deteriorates
Solution Approach 1:
The training data is segmented into two categories processed in sequence: text-based query-document pairs for Phase 1, and interaction-enriched data for Phase 2. This segmentation ensures that the semantic understanding foundation is built on high-quality text semantics without the noise of interaction biases, while still utilizing all available data across both phases for comprehensive model optimization.
Solution Approach 2:
Text-based semantic training is performed preliminarily before introducing interaction data, ensuring that the model's semantic understanding capabilities are established on pure linguistic foundations. This preliminary text-focused training preserves measurement precision for semantic tasks, while subsequent phases layer on interaction-based personalization without compromising the core semantic understanding.
3Adaptability or versatility
If transformer models are trained with user interaction data from the beginning, then personalized ranking capability is improved, but the complexity of training and computation increases
Solution Approach 1:
The training complexity is segmented across two phases with different computational requirements. Phase 1 uses simpler text-based training without interaction data, requiring less computational complexity. Phase 2 then introduces user interaction data and personalized ranking objectives building on the Phase 1 foundation. This segmentation manages device complexity by staggering the introduction of computational demands while still achieving personalized ranking capability.
Solution Approach 2:
The model performs preliminary training on text-based data to establish baseline capabilities before introducing the complexity of interaction-based personalized ranking. This preliminary action reduces overall training complexity by preparing the model architecture and parameters in advance, making the subsequent personalized ranking training more efficient and manageable despite the increased adaptability requirements.
4Loss of energy
If traditional relevance determination methods are used, then computational resources are conserved, but the accuracy of search result ranking deteriorates
Solution Approach 1:
The system performs preliminary training on a condensed text-based dataset in Phase 1, establishing efficient baseline representations. This preliminary action creates optimized initial parameters that reduce the computational burden of subsequent personalized ranking computations. The two-phase approach conserves computational resources by avoiding redundant processing while maintaining high ranking accuracy through the staged introduction of interaction data.
Data Source
AI summary
A method and system for training a machine-learning algorithm (MLA) to rank digital documents at an online search platform. The method comprises training the MLA in a first phase for determining past user interactions of a given user with past digital documents based on a first set of training objects including the past digital documents generated by the online search platform in response to the given user having submitted thereto respective past queries. The method further comprises training the MLA in a second phase to determine respective likelihood values of the given user interacting with in-use digital documents based on a second set of training objects including only those past digital documents with which the given user has interacted and respective past queries associated therewith. The MLA may include a Transformer-based learning model, such as a BERT model.