| dc.description.abstract | The translation task has always been a topic of great concern in the field of natural language, and even though many previous studies of well-known translation tasks have contributed to translation performance and effectiveness, they have been limited to dealing with single pair or ＂English-centric＂ tasks. Recently, more and more research has been conducted on multilingual, non-English-centric, machine translation systems.
Current multilingual machine translation methods still suffer from data imbalance in training. Recent studies have used contrastive learning methods to reduce the representation gap between different languages to improve the performance of multilingual machine translation. On the other hand, a study in the field of computer vision has questioned the well- known contrastive learning framework of negative sampling, arguing that unconditional selection of negative samples can cause unstable learning, and therefore hypothesizing that conditional selection of negative samples can help improve the performance of contrastive learning and prove its effectiveness. Since there is no task to discuss the problem of negative sampling in contrastive learning on multilingual translation tasks, this study aims to discuss the method of setting conditional negative samples in contrastive learning.
From the experimental results, we found that our proposed conditional contrastive learning method was not as effective as we expected in supervised directional translation. Insufficient amount of training data makes the contrastive learning method of self-supervised learning ineffective in improving multilingual machine translation, making it difficult for multi-task learning models to outperform supervised learning models in translation.
We further analyze the shared sentence representations learned using our proposed model. We also visualize and compare these shared representations with m-Transformer and contrastive learning models for single-task learning, and demonstrate that our shared representations can be effective in learning shared sentence representations across languages. We obtained similar results to the contrastive learning model, and we believe that our advantage lies in the fact that the learning targets are sampled beforehand instead of extracting all objects for learning as in the previous approach, and that using fewer learning objects reduces the learning instability problem. | en_US |