Contextual Density Ratio Biasing for Sequence-to-Sequence Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
End-to-end (E2E) sequence-to-sequence processing systems, such as automated speech recognition (ASR) systems, face difficulties in dynamically adapting to specific contextual profiles for better processing of entities like named entities or special terms due to their internal language models, which make it challenging to integrate external language models effectively.
Innovation Solution
A method that involves identifying specialized entities within candidate output sequences, applying a first scoring methodology based on an internal language model, and a second scoring methodology based on an external language model, with the prediction scores for these entities being modified by combining the scores to enhance recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If end-to-end sequence-to-sequence processing systems are used, then processing performance is improved, but adaptability to specific contextual profiles deteriorates
Solution Approach 1:
The patent segments the scoring process into two distinct methodologies: a first scoring methodology for general candidate output sequences and a second scoring methodology for specialized entities. This segmentation allows the system to maintain strong general processing performance while separately optimizing for specialized entity recognition, thereby resolving the contradiction between overall performance and contextual adaptability.
Solution Approach 2:
The patent applies different scoring qualities to different parts of the output sequence. General output sequences use the first scoring methodology based on internal language models, while specialized entities use the second scoring methodology based on external language models. This local differentiation enables the system to optimize each component according to its specific requirements, improving both overall performance and contextual adaptability.
2Productivity
If internal language models are used in E2E systems, then processing efficiency is improved, but integration of external language models deteriorates
Solution Approach 1:
The patent merges the internal language model scoring and external language model scoring into a unified output sequence generation process. By combining both scoring methodologies and their respective predictions, the system achieves both processing efficiency through the internal model and adaptability through the external model integration, resolving the contradiction between efficiency and integration capability.
Solution Approach 2:
The patent creates a universal scoring framework that can accommodate both internal and external language models. The system uses the first scoring methodology for general cases and the second scoring methodology for specialized entities, making the system multi-functional and capable of integrating external models without sacrificing internal processing efficiency.
3Measurement precision
If prediction scores for specialized entities are modified using external language models, then recognition accuracy for specialized entities is improved, but system complexity increases
Solution Approach 1:
The patent applies partial action by selectively applying the second scoring methodology only to specialized entities rather than to all output sequences. This partial application improves recognition accuracy for specialized entities while minimizing the increase in system complexity, as the external language model integration is confined to specific portions of the processing pipeline.
Solution Approach 2:
The patent introduces an intermediary scoring adjustment mechanism that mediates between the internal language model predictions and external language model predictions. This intermediary layer combines the two scoring methodologies in a controlled manner, improving specialized entity recognition accuracy while managing system complexity through a structured integration approach.
Data Source
AI summary
A method, computer program product, and computer system for processing one or more portions of an input sequence to generate one or more candidate output sequences, thus defining a plurality of prediction scores for the candidate output sequences. One or more specialized entities may be identified from the candidate output sequences. A first scoring methodology may be applied on the candidate output sequences based upon the portions of the input sequence, thus defining a first set of prediction scores for the one or more candidate output sequences. A second scoring methodology may be applied on the specialized entities from the candidate output sequences based upon the portions of the input sequence, thus defining a second set of prediction scores for the specialized entities. The plurality of predictions scores for the specialized entities may be at least partially modified based upon the first set and the second set of prediction scores.