Text Compression Dictionary Mapping for Faster Neural Translation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current translation techniques using neural networks face processing speed deterioration due to frequent vector conversion for idioms and inflections, and struggle with optimizing vector allocation for polysemous words and phrases, leading to inefficiencies in machine learning and translation processes.
Innovation Solution
The method involves generating a compression dictionary that associates word frequencies with meaning, using static and dynamic codes to efficiently convert text data into neural network codes (NNCs), which are then converted into vector values, reducing the need for repeated decompression and morphological analysis, and enhancing processing speed and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If neural network translation techniques are used with frequent vector conversion for idioms and inflections, then translation capability is improved, but processing speed deteriorates
Solution Approach 1:
The patent pre-generates and stores conversion dictionaries that map compressed codes to neural network codes (NNCs) and vector values before translation processing. When translating, the system directly retrieves pre-computed vectors from these dictionaries using compressed codes as keys, eliminating the need for repeated decompression and morphological analysis during actual translation, thus resolving the contradiction between translation capability and processing speed
Solution Approach 2:
The patent creates compressed versions of text data using compression algorithms and stores corresponding pre-computed neural network codes and vector values in conversion dictionaries. Instead of processing original text repeatedly, the system works with compact compressed codes that directly index into pre-computed vector representations, maintaining translation quality while dramatically improving processing speed
2Measurement precision
If decompression and morphological analysis are performed repeatedly for text data, then accurate vector generation is achieved, but processing efficiency decreases
Solution Approach 1:
The patent performs decompression and morphological analysis in advance to generate accurate vector representations, then stores these results in conversion dictionaries indexed by compressed codes. During actual translation, the system directly retrieves pre-computed vectors without repeating decompression or morphological analysis, thus maintaining accuracy while dramatically improving processing efficiency
Solution Approach 2:
The conversion dictionaries serve as self-service lookup tables that contain all necessary vector information indexed by compressed codes. The translation system can directly query these dictionaries without needing to perform repeated decompression or morphological analysis, making the process self-sufficient and highly efficient
Data Source
AI summary
A computer: acquires, from a compression dictionary that associates each of codes having a length according to a frequency of appearance of a set of a word and a word meaning of the word with that set, the set of any word and that word meaning, and one of the codes associated with the set of the any word and that word meaning; selects, from among a plurality of fixed-length codes stored in the memory with a same length in association with the set of the word and that word meaning, one of the fixed-length codes associated with the set of the any word and that word meaning; generates a conversion dictionary that associates the selected one of the fixed-length codes with the acquired one of the codes; and specifies, by the conversion dictionary, the individual fixed-length codes associated with each of the codes contained in compressed data.