Two-Tower Multi-Task Model for Semantic Retrieval and Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing e-commerce systems face challenges in efficiently training multiple models for tasks such as semantic retrieval, embedding, and classification due to high energy consumption, maintenance costs, and complex relationships between queries and items, leading to suboptimal performance and hindered AI development.
Innovation Solution
A multi-task machine learning model utilizing a two-tower structure is trained with a bi-directional transformer architecture to optimize query and item embeddings, enabling simultaneous support for tasks like semantic item retrieval, query classification, and clustering, while reducing training time and energy consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If individual domain-specific models are built for different tasks, then task-specific performance is improved, but training time and energy consumption increase significantly
Solution Approach 1:
The patent implements a universal multi-task machine learning model that can perform multiple e-commerce tasks (semantic retrieval, embedding, classification) simultaneously. The model uses shared neural network layers to process different task types, allowing a single model to replace multiple individual domain-specific models, thereby reducing training time while maintaining task-specific performance through task-specific output heads.
Solution Approach 2:
The patent combines multiple individual models into a unified multi-task model by merging their functional capabilities. The model integrates task-specific processing for semantic retrieval, embedding, and classification into a single architecture with shared representations, eliminating the need to train separate models for each task and significantly reducing overall training time.
2Measurement precision
If individual domain-specific models are built for different tasks, then task-specific performance is improved, but energy consumption increases significantly
Solution Approach 1:
The universal multi-task model reduces energy consumption by performing multiple tasks through a single trained model rather than requiring separate training runs for each task. The shared neural network layers process multiple task types simultaneously, eliminating redundant computational work and reducing overall energy consumption while maintaining the ability to perform task-specific operations.
Solution Approach 2:
By merging multiple individual models into one unified model, the system eliminates the energy overhead of training separate models for each task. The combined model shares computational resources and learned representations across tasks, reducing total energy consumption while preserving task-specific performance through specialized output layers.
3Adaptability or versatility
If multiple individual models are maintained for different tasks, then task coverage is improved, but maintenance cost increases
Solution Approach 1:
The multi-task model provides comprehensive task coverage (semantic retrieval, embedding, classification) through a single universal model, eliminating the need to maintain multiple separate models. This unified approach reduces maintenance costs by requiring only one model to train, update, and deploy, while still supporting diverse e-commerce tasks through its multi-functional architecture.
Solution Approach 2:
The patent merges multiple task-specific models into a single maintainable unit. Instead of separately managing multiple models with different training pipelines and update schedules, the system maintains one unified model that handles all tasks, significantly reducing the complexity and cost of model maintenance while preserving broad task coverage.
4Measurement precision
If multiple individual models are trained for different tasks, then task specialization is improved, but learning curve complexity increases
Solution Approach 1:
The multi-task model achieves task specialization for semantic retrieval, embedding, and classification within a single unified framework. Practitioners only need to learn one model architecture and training process rather than multiple separate models, reducing the learning curve complexity while maintaining the ability to perform specialized tasks through the model's multi-functional capabilities.
Solution Approach 2:
By combining multiple task-specific models into one, the patent reduces the complexity of the learning curve. Instead of requiring expertise in training and managing multiple separate models with different requirements, the system uses a single unified model that handles all tasks, making it easier for practitioners to learn and deploy while preserving task specialization through the model's architecture.
Data Source
AI summary
Systems and methods for generating a machine learning model to support multiple tasks for semantic retrieval, embedding and classification are disclosed. In some embodiments, a disclosed method includes: obtaining a training data set generated based on search related data and item related data associated with a website; training, based on the training data set, a machine learning model using a two-tower structure to generate an optimized set of model parameters, wherein the optimized set of model parameters minimizes a total loss function computed based on at least one of: a query classification loss, an item classification loss, and a similarity loss; and transmitting the trained machine learning model to be utilized for at least one of the following tasks: semantic item retrieval, query classification, item classification, query clustering, and item clustering.