Fusion of Improved BERT and BiLSTM Models to Improve English-Chinese Terminology Translation

Abstract

To address weak domain adaptability and insufficient modeling of long-range contextual dependencies in specialized English-Chinese terminology translation, this paper proposes a neural translation model integrating improved BERT and BiLSTM. The model first pre-trains BERT on the public WMT corpus and domain-specific corpora, and introduces term boundary markers to enhance input representation, improving recognition of specialized terminology and domain adaptability. A gated-attention BiLSTM encoder is then designed to strengthen local contextual awareness and sequence-structure modeling, especially around term boundaries, thereby complementing BERT’s global representation and capturing the semantic context of terms more accurately. A feature-fusion module integrates multi-layer representations from BERT and BiLSTM, and an attention-driven decoder generates the final translation. Experimental results show that the proposed model outperforms advanced baselines across key metrics. Terminology translation accuracy reaches 88.1%, domain consistency reaches 92.0%, term-boundary F1-score reaches 87.6%, and context sensitivity reaches 86.5%, verifying the effectiveness of the method in improving the accuracy and robustness of specialized terminology translation.