Offline Neural Network Translation Model with Vocabulary Rollback Mechanism
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional offline translation models executed locally on devices suffer from lower capacity factors, leading to reduced translation quality and a higher propensity to generate neologisms due to limited computational resources, compared to remotely executed models.
Innovation Solution
A local translation model is built using a Neural Network with an encoder and decoder portion, incorporating an attention mechanism, which allows for iterative 'rolling back' to previous states to generate alternative translation candidates, ensuring that the final translation adheres to pre-determined vocabulary rules and reduces neologism generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If a translation model is executed locally on a user device, then offline translation capability is provided, but translation quality deteriorates due to lower capacity factor
Solution Approach 1:
The translation model is segmented into an encoder portion and a decoder portion. The encoder processes the source language input and generates intermediate representations, while the decoder generates target language output. This segmentation allows the model to be optimized for local execution while maintaining translation quality through specialized processing in each portion.
Solution Approach 2:
The translation model is trained in advance on a server with high computational resources before being deployed to the local device. The training phase performs preliminary learning of translation patterns and vocabulary, enabling the compact local model to achieve better translation quality without requiring extensive local computational resources during actual translation operations.
2Ease of operation
If a translation model is executed locally on a user device, then offline translation capability is provided, but neologism generation increases due to lower capacity factor
Solution Approach 1:
A vocabulary verification mechanism provides feedback during the decoding process. The system checks generated candidate translations against a pre-stored vocabulary database and triggers roll-back when neologisms are detected. This feedback loop ensures that only valid words from the target vocabulary are produced, eliminating the harmful effect of neologism generation while maintaining offline capability.
Solution Approach 2:
A vocabulary database is pre-loaded and stored in the local device before translation operations begin. This preliminary preparation allows the verification mechanism to quickly check generated candidates against known valid words, preventing neologism generation without requiring extensive computational resources during actual translation.
3Manufacturing precision
If the decoder generates multiple candidate translations, then translation accuracy improves, but computational resources increase
Solution Approach 1:
The system generates multiple candidate translations partially, focusing only on the most probable candidates rather than exhaustively exploring all possible translations. The decoder generates a limited set of top candidates and verifies them against the vocabulary, achieving improved accuracy without the full computational cost of generating and evaluating all possible translation combinations.
Solution Approach 2:
The candidate generation and verification process is segmented into distinct phases: candidate generation by the decoder, vocabulary verification by the verification mechanism, and selective acceptance or roll-back. This segmentation allows efficient resource management by only performing full verification on promising candidates rather than all possible translations.
Data Source
AI summary
Methods and electronic devices for executing offline translation of a source word into a target word via a Neural Network an encoder and a decoder. The method includes splitting the source word into input tokens, generating vector representations for input tokens, and generating a first sequence of output tokens representative of a first candidate word. In response to the first candidate word not respecting at least one pre-determined rule, the method includes triggering the decoder to generate a second sequence of output tokens having a different at least one last output token than at least one last output token of the first sequence. The second sequence is representative of a second candidate word. In response to the second candidate word respecting the at least one pre-determined rule, the method includes determining that the second candidate word is the target word.