Sequence Conversion Using Coverage Vector for Faithful Translation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sequence conversion methods in natural language processing, such as sequence-to-sequence learning, fail to accurately transmit information from the source sequence to the target sequence due to the RNN's reliance on a target context vector, leading to a lack of faithfulness in the predicted target sequence.
Innovation Solution
A computer-implemented sequence conversion method that obtains a weight vector using the reference and source context vectors, weights these vectors, and predicts a target context vector using a weighted source and reference context vector, dynamically determining weights to improve faithfulness and smoothness of the target sequence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the RNN uses a target context vector as a main reference and a source context vector only as an extra input, then the decoding process can be simplified and operated efficiently, but the information corresponding to the source context vector cannot be correctly transmitted to the target context vector, resulting in loss of information
Solution Approach 1:
The patent introduces a coverage vector as an intermediary element that tracks the correspondence between source and target sequences. This coverage vector acts as a mediator to ensure that information from the source context vector is properly transmitted to the target context vector, resolving the information loss problem while maintaining decoding efficiency
Solution Approach 2:
The patent implements a feedback mechanism where the coverage vector continuously monitors and provides information about which source words have been translated. This feedback loop ensures that the target context vector receives accurate information from the source context vector, preventing information loss while maintaining efficient decoding operations
2Ease of operation
If the target sequence is generated by relying mainly on the target context vector from previous predictions, then the generation process is streamlined, but the target sequence cannot faithfully reflect information about the source sequence, resulting in low faithfulness
Solution Approach 1:
The coverage vector serves as an intermediary that bridges the target context vector and source sequence information. It ensures that the simplified generation process still faithfully reflects source sequence information by mediating the information flow between these components
Solution Approach 2:
The patent creates a composite context representation by combining the target context vector with coverage vector information. This composite structure maintains the simplicity of generation while ensuring faithfulness to the source sequence, as it integrates multiple information sources into a unified representation
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The present invention discloses a sequence conversion method and apparatus, and the sequence conversion method includes: obtaining a source sequence; obtaining a reference context vector; converting the source sequence into a source context vector; obtaining a weight vector according to the reference context vector and the source context vector; weighting the source context vector and the reference context vector by using the weight vector; predicting, by using a weighted source context vector and a weighted reference context vector, a target context vector corresponding to the source context vector; and obtaining a target sequence according to the target context vector. According to the present invention, faithfulness of the target sequence to the source sequence during sequence conversion can be improved.