Multi-dimensional Style Transfer via Cascaded Language Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing authoring methods are unable to simultaneously transfer input text into multiple style dimensions, such as formality and sentiment, leading to unsatisfactory results when regenerating text across various style combinations.
Innovation Solution
A system utilizing a cascaded language model trained with multiple discriminator models for each target style, allowing for the transfer of a sentence into a multidimensional style combination by minimizing a loss function that includes terms for the cascaded language model and multiple discriminator models, enabling simultaneous transfer across multiple styles with higher accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing authoring methods are used to transfer text style, then text can be regenerated in a single style dimension, but the method is unable to simultaneously transfer text into multiple style dimensions
Solution Approach 1:
The patent divides the style transfer task into multiple independent discriminator models, each responsible for a specific style dimension (e.g., formality, sentiment). Each discriminator independently evaluates one aspect of the transferred text, allowing the system to handle multiple style dimensions simultaneously while keeping each component relatively simple and manageable.
Solution Approach 2:
The patent transitions from single-dimension style transfer to multi-dimension style transfer by introducing multiple discriminator models that operate in parallel across different style dimensions. This dimensional expansion allows the system to evaluate and transfer multiple style attributes simultaneously, moving from a one-dimensional to a multi-dimensional style transfer framework.
2Measurement precision
If existing authoring methods transfer text style, then processing time is reduced compared to manual rewriting, but accuracy across multiple styles deteriorates
Solution Approach 1:
The patent employs pre-trained discriminator models that have been trained in advance on specific style dimensions. These pre-trained discriminators are then fine-tuned or combined for the specific multi-dimensional style transfer task, eliminating the need to train from scratch and significantly reducing training time while maintaining high accuracy across multiple style dimensions.
Solution Approach 2:
The patent adjusts training parameters and loss function weights to optimize the balance between training time and accuracy. By modifying parameters such as learning rates, batch sizes, and loss function coefficients, the system achieves high style transfer accuracy across multiple dimensions while controlling training time within acceptable limits.
3Reliability
If existing authoring methods are used, then implementation is simpler, but content preservation across multiple styles deteriorates
Solution Approach 1:
The patent implements a feedback mechanism where multiple discriminator models evaluate the transferred text and provide feedback signals. Each discriminator assesses a specific style dimension and provides guidance back to the generator model, ensuring that the transferred text maintains content integrity while achieving the desired multi-dimensional style transformation. This feedback loop continuously refines the output to preserve content across style changes.
Solution Approach 2:
The patent creates a composite model architecture that integrates multiple discriminator models with different style evaluation capabilities. This composite structure combines the strengths of individual discriminators specialized in different style dimensions, enabling the system to preserve content while achieving accurate multi-dimensional style transfer that no single discriminator could accomplish alone.
Data Source
AI summary
In some embodiments, a style transfer computing system receives, from a computing device, an input text and a request to transfer the input text to a target style combination including a set of target styles. The system applies a style transfer language model associated with the target style combination to the input text to generate a transferred text in the target style combination. The style transfer language model comprises a cascaded language model configured to generate the transferred text. The cascaded language model is trained using a set of discriminator models corresponding to the set of target styles. The system provides, to the computing device, the transferred text.