Machine Translation Vector Retrieval for Accuracy and Overhead
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional Neural Machine Translation (NMT) systems face limitations in translation efficiency due to the high calculation overhead of Hierarchical Recurrent Neural Network (HRNN) modules, which restricts the summary of previous sentence translations to a few sentences, leading to a limited improvement in translation effect.
Innovation Solution
The method involves obtaining a source representation vector for a word fragment, querying historical translation information to obtain a historical translation vector, and using this vector to translate the word fragment, thereby considering the comprehensive translation situation in the historical process, improving translation accuracy by leveraging past translation scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If HRNN network is used to summarize translation results of previous sentences, then translation accuracy is improved, but calculation overhead increases significantly
Solution Approach 1:
The patent extracts only the necessary historical translation information (translation results of previous sentences) from the full context, storing them as separate vector representations. This allows the system to access relevant historical data without processing the entire translation history, reducing calculation overhead while maintaining translation accuracy.
Solution Approach 2:
The patent performs preliminary processing by pre-calculating and storing translation result vectors for previous sentences in memory. When translating the current sentence, the system directly retrieves these pre-computed vectors rather than re-calculating them, significantly reducing the computational burden during the actual translation process.
2Productivity
If only a few previous sentences are summarized to ensure translation efficiency, then calculation overhead is reduced, but translation effect improvement is limited
Solution Approach 1:
The patent transitions from a time-based limitation (summarizing only the last 3 sentences) to a memory-based approach where all previous translation results can be stored and accessed. By organizing historical translations as searchable vector representations in memory, the system can efficiently access any number of previous sentences without proportionally increasing computational complexity.
Solution Approach 2:
The patent creates vector representations (copies) of previous translation results that can be stored in memory and retrieved as needed. These vector copies contain the essential information from historical translations without requiring the system to re-process the original text, enabling efficient access to extensive historical context.
3Measurement precision
If HRNN network processes more previous sentences, then translation accuracy improves, but translation efficiency decreases
Solution Approach 1:
The patent segments the historical translation context into individual sentence-level vector representations. Each previous sentence's translation result is stored as a separate vector in memory, allowing the system to selectively access and process only the relevant historical information needed for the current translation, rather than processing all historical data uniformly.
Solution Approach 2:
The patent introduces a memory structure as an intermediary between the historical translation results and the current translation process. This memory acts as a buffer that stores vector representations of previous translations, enabling the system to access extensive historical context without the computational overhead of processing all that data through the HRNN network.
Data Source
AI summary
A machine translation method and an apparatus are provided. The method includes: obtaining, when translating a word fragment from a first language into a second language, a source representation vector of the word fragment. The source representation vector of the word fragment represents the word fragment in the first language. The method also includes obtaining a historical translation vector of the word fragment by querying historical translation information according to the source representation vector of the word fragment. The historical translation vector of the word fragment represents a historical translation situation corresponding to the word fragment. The method further includes translating the word fragment according to the historical translation vector of the word fragment.