Scholarlink

<p>Assessing student translations in real classrooms still leans heavily on human judgment, which can vary across raters, is time-consuming, and scales poorly. With the rapid advancement of natural language processing (NLP), automatic assessment models have shown potential to enhance objectivity and consistency in translation evaluation. This paper proposes BERT-SVM EduScore, a hybrid assessment model designed for English translation education. The model couples a BERT-base encoder adapted with domain-/task-specific training and a contrastive objective with compact linguistic and alignment features, and feeds the resulting representations into margin-based Support Vector Machines (SVM) heads with monotonic calibration to produce holistic and rubric-aligned sub-scores. We conduct experiments on the English–Chinese portion of the MLQE-PE machine-translation quality estimation dataset, which we use as a proxy for sentence-level scoring in classroom-like settings. BERT-SVM EduScore is compared against string-based metrics (BLEU, METEOR, TER, chrF++), embedding-based semantic metrics (YiSi-1, MoverScore, BERTScore), and learned evaluators (BLEURT, COMET, PRISM, BARTScore, TransQuest) using Quadratic Weighted Kappa (QWK), mean absolute error (MAE), Pearson’s r, and runtime. On MLQE-PE it achieves QWK 0.76, MAE 0.12, and r 0.84, improving QWK over COMET by + 0.08 and over TransQuest by + 0.05 while running at about 22.5 ms per sentence (≈ 44 sentences per second) on commodity GPU hardware. Ablation studies show that domain-adaptive pretraining yields the largest gains, contrastive learning provides additional improvements, and metric distillation contributes smaller but consistent benefits. These quantitative results suggest that the proposed model can serve as a technically feasible component for translation education, offering calibrated scores and lightweight diagnostic signals under realistic latency constraints; validation on real course assignments and with teachers and students in the loop is left for future work.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A hybrid intelligent assessment model for English translation education with improved BERT and SVM

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月19日

查看全文
  • Chuan Lin

摘要

Assessing student translations in real classrooms still leans heavily on human judgment, which can vary across raters, is time-consuming, and scales poorly. With the rapid advancement of natural language processing (NLP), automatic assessment models have shown potential to enhance objectivity and consistency in translation evaluation. This paper proposes BERT-SVM EduScore, a hybrid assessment model designed for English translation education. The model couples a BERT-base encoder adapted with domain-/task-specific training and a contrastive objective with compact linguistic and alignment features, and feeds the resulting representations into margin-based Support Vector Machines (SVM) heads with monotonic calibration to produce holistic and rubric-aligned sub-scores. We conduct experiments on the English–Chinese portion of the MLQE-PE machine-translation quality estimation dataset, which we use as a proxy for sentence-level scoring in classroom-like settings. BERT-SVM EduScore is compared against string-based metrics (BLEU, METEOR, TER, chrF++), embedding-based semantic metrics (YiSi-1, MoverScore, BERTScore), and learned evaluators (BLEURT, COMET, PRISM, BARTScore, TransQuest) using Quadratic Weighted Kappa (QWK), mean absolute error (MAE), Pearson’s r, and runtime. On MLQE-PE it achieves QWK 0.76, MAE 0.12, and r 0.84, improving QWK over COMET by + 0.08 and over TransQuest by + 0.05 while running at about 22.5 ms per sentence (≈ 44 sentences per second) on commodity GPU hardware. Ablation studies show that domain-adaptive pretraining yields the largest gains, contrastive learning provides additional improvements, and metric distillation contributes smaller but consistent benefits. These quantitative results suggest that the proposed model can serve as a technically feasible component for translation education, offering calibrated scores and lightweight diagnostic signals under realistic latency constraints; validation on real course assignments and with teachers and students in the loop is left for future work.