Encoder Vector Validation for Neural Machine Translation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural machine translation technologies face delays and resource inefficiencies when translating text from one language to another, as they require acquiring multiple translation texts and then processing them to identify and correct abnormal translations, leading to slowed operation speeds and degraded hardware resource utilization.
Innovation Solution
An electronic device and method that utilize an encoding vector acquired through an encoder to determine the quality of a translation, transmitting it to an external decoder only if it meets a certain threshold, or replacing it with a similar vector if it doesn't, thereby bypassing the need for initial decoder operation and improving translation efficiency and quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If translation texts are acquired first and then processed to identify and correct abnormal translations, then translation quality can be improved, but operation speed is delayed and hardware resource utilization is degraded
Solution Approach 1:
The patent applies preliminary action by acquiring and validating encoding vectors before decoder operations. The system validates encoding vectors in advance using a validation model to determine whether they represent normal or abnormal translations. This preliminary validation allows the system to prepare appropriate replacement encoding vectors beforehand, avoiding the need for post-processing decoder operations and thereby improving both translation quality and operation speed.
2Measurement precision
If multiple translation texts are acquired and processed to correct abnormal translations, then translation accuracy improves, but hardware resource utilization is degraded
Solution Approach 1:
The patent extracts and validates only the essential encoding vectors required for translation without processing multiple complete translation texts. By focusing validation efforts on encoding vectors alone rather than entire translation outputs, the system achieves high translation accuracy while minimizing hardware resource consumption. The validation model specifically targets encoding vector quality metrics, extracting only the necessary information needed for accurate translation.
3Reliability
If conventional translation processing is performed with multiple texts and post-processing, then translation quality can be maintained, but operation efficiency is reduced
Solution Approach 1:
The system performs preliminary validation of encoding vectors before they are used in decoder operations. By validating encoding vectors in advance and preparing replacement vectors beforehand, the system eliminates time-consuming post-processing steps. This preliminary action ensures translation quality is maintained while significantly improving operation efficiency by avoiding delays associated with post-processing multiple translation texts.
Data Source
AI summary
An electronic device and a controlling method of the electronic device are disclosed. In particular, the electronic device according to the disclosure includes a communicator, a memory configured to store information on an encoder and learning data for learning of the encoder, and a processor configured to, based on acquiring information on an original text of a first language, input the information on the original text into the encoder, and acquire a first encoding vector indicating semantic information included in the original text, input the first encoding vector into a discriminator, and acquire a probability value indicating a probability that the first encoding vector would correspond to a normal translation, based on the probability value being greater than or equal to a predetermined first threshold value, control the communicator to transmit the first encoding vector to an external device including a decoder for acquiring a translation text of a second language corresponding to the original text, and based on the probability value being smaller than the first threshold value, identify a second encoding vector having the highest similarity value indicating similarity to the first encoding vector among encoding vectors included in the learning data, and control the communicator to transmit the identified second encoding vector to the external device.