Machine translation quality estimation (QE) refers to the task of estimating the quality of machine translation output without relying on reference translations. Existing QE systems perform well in general domains, but their performance declines in specialized domains that contain a large number of domain-specific terms. This is because current methods primarily focus on assessing the semantic similarity between the source sentence and the target translation, lacking sensitivity to translation errors involving specialized terminology. To address this problem, we propose a domain-specific translation quality estimation method that incorporates cross-lingual fine-grained dual alignment features. First, we design prompt templates to guide GPT in generating bilingual term alignment information, thereby obtaining bilingual term alignment representations. Next, we optimize the encoder of the word alignment model to meet the requirements of the QE task and extract vector representations that contain fine-grained word alignment information from the alignment layer. Finally, we incorporate the dual alignment features as explicit features into the estimator of the QE model, combining them with neural features. These features undergo nonlinear transformations through a feed-forward neural network to generate the final prediction score. Experimental results demonstrate the effectiveness of our method. Compared to state-of-the-art methods, the main metric of Spearman’s correlation coefficient improves by 5.69%, while the auxiliary metrics of Pearson’s correlation coefficient and Kendall’s correlation coefficient improve by 5.8% and 4.96%, respectively.