SE-Enhancer: Low-Resource Machine Translation Based on Enhanced SimCSE and Layer Fusion

Dongsheng Wang, Qile Bo, Fei Wang, Liming Wang, Kun Tang

Abstract: Low-resource machine translation holds significant practical importance for the translation of small languages. Currently, the primary challenge in low-resource machine translation is the scarcity of bilingual parallel corpora. To address this issue, this paper proposes a SE-Enhancer model based on enhanced SimCSE. We first introduced an embed-fusion module to integrate word embeddings and sentence embeddings, thereby enriching the feature representation of source sentences. Then, a layer fusion module based on feedforward neural networks and self-attention is integrated into the model to reduce information loss by integrating multi-layer features, enhancing the decoder’s translation capabilities and overall model performance. The experimental results demonstrate that SE-Enhancer achieves BLEU score improvements of 1.24, 1.37, and 1.11 over the Transformer baseline model on three common low-resource machine translation tasks from the IWSLT dataset. Length analysis further demonstrates that the model excels at capturing long-distance dependencies and exhibits strong generalization capabilities. In practical applications, this model can be utilized in low-resource fields such as legal, healthcare, and education translation.

External IDs:dblp:journals/access/WangBWWT25