Bidirectional Translation Model Training via Pseudo-Parallel Corpus
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In machine translation, the scarcity of bilingual parallel corpuses for minority languages hinders the effective training of machine translation models due to high acquisition costs and difficulties in obtaining high-quality reverse translation models.
Innovation Solution
A bidirectional translation model is trained using a method that includes N cycles of forward and reverse translation processes, with the convergence of forward and reverse translation similarities determining the completion of training, and utilizes a reconstructor and differentiable sampling functions to optimize both translation capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If bilingual parallel corpuses are used for training machine translation models, then translation model performance is improved, but acquisition cost and difficulty increase significantly
Solution Approach 1:
The patent creates pseudo-parallel corpuses by copying and transforming monolingual data through bidirectional translation. The reconstructor model generates pseudo-target texts from source texts, and pseudo-parallel data is constructed by pairing source texts with their reconstructed versions, eliminating the need for expensive human-annotated bilingual parallel corpuses
Solution Approach 2:
The system uses the translation model itself to generate training data through self-reconstruction. The bidirectional translation process where target texts are translated back to source texts and compared with original source texts creates a self-supervised learning mechanism that eliminates dependency on external bilingual resources
2Measurement precision
If high-quality reverse translation models are obtained, then training data quality is improved, but acquisition difficulty increases
Solution Approach 1:
The patent merges the forward translation model and reverse translation model into a single bidirectional training framework. Both models are trained simultaneously using the same pseudo-parallel corpus, sharing parameters and training processes, which eliminates the need to separately acquire high-quality reverse translation models
Solution Approach 2:
Instead of acquiring reverse translation models through traditional methods, the patent inverts the approach by using the forward translation model to generate pseudo-target texts, then using these to train the reverse model. The reconstruction quality is measured by back-translation similarity rather than by comparing with reference translations
Data Source
AI summary
A method includes: a bidirectional translation model to be trained and training data are acquired, the training data including a source corpus and a target corpus corresponding to the source corpus; the bidirectional translation model is trained for N cycles, each cycle of training including a forward translation process of translating the source corpus into a pseudo target corpus and a reverse translation process of translating the pseudo target corpus into a pseudo source corpus and N being a positive integer greater than 1; a forward translation similarity and a reverse translation similarity are acquired; and when a sum of the forward translation similarity and the reverse translation similarity converges, it is determined that training of the bidirectional translation model is completed, where the training completed bidirectional translation model is used to perform translating.