Beam Search Text Generation with Decoded-Word De-Duplication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing text generation methods using Beam Search algorithms are inefficient and inaccurate due to the lack of de-duplication of decoded-word information, leading to a large number of duplicated texts and reduced efficiency and accuracy in generating text themes.
Innovation Solution
A text generation method that involves determining a decoded-word information group set, de-duplicating it to form a candidate-word information group set, selecting candidate-word information that meets a target condition, and splicing it with a historical target-word information sequence based on a preset-word list to generate a target text.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If Beam Search algorithm is used for text generation, then text generation capability is provided, but efficiency and accuracy are low due to duplicated texts
Solution Approach 1:
The patent extracts and removes duplicated text information from the decoded-word information group set to create a candidate-word information group set. This extraction process eliminates redundant entries while preserving unique text themes, directly addressing the efficiency and accuracy problems caused by duplicated texts in Beam Search decoding.
Solution Approach 2:
The patent changes the parameter of information representation by transforming decoded-word information into candidate-word information through deduplication. This parameter transformation optimizes the data structure for subsequent selection and splicing operations, improving overall generation efficiency and accuracy.
2Adaptability or versatility
If duplicated texts are included in decoding process, then comprehensive text coverage is achieved, but processing time increases and efficiency decreases
Solution Approach 1:
The patent extracts only unique text themes from the decoded-word information by removing duplicates. This selective extraction maintains comprehensive text coverage of unique themes while eliminating redundant processing of duplicated texts, thereby reducing processing time and improving efficiency.
Solution Approach 2:
The patent performs deduplication as a preliminary action before the main selection and splicing processes. By preparing the candidate-word information group set in advance with duplicates removed, subsequent operations work with optimized data, reducing overall processing time while maintaining text coverage.
3Loss of information
If all decoded-word information is processed, then complete information is maintained, but computational complexity increases
Solution Approach 1:
The patent extracts and retains only unique text theme information while removing duplicated entries. This extraction maintains information completeness regarding unique themes while significantly reducing the volume of data requiring processing, thereby lowering computational complexity.
Solution Approach 2:
The patent changes the parameter of data volume by transforming the decoded-word information group set into a smaller candidate-word information group set through deduplication. This parameter change reduces computational complexity while preserving the essential information completeness needed for accurate text generation.
Data Source
AI summary
Text generation including de-duplication of decoded word information to splice target word information into an information sequence. In a specific scheme, a decoded-word information group set is determined on the basis of a text to be processed; the decoded-word information group set is de-duplicated to generate a candidate-word information group set; candidate-word information that meets a target condition as target-word information is selected from each candidate-word information group in the candidate-word information group set to obtain a target-word information set; if the target-word information meets a convergence condition is determined, the target-word information with a historical target-word information sequence corresponding to the target-word information is spliced on the basis of a preset-word list to generate a target text.