Scholarlink

<p>Idiomatic expressions remain difficult for neural machine translation because their meanings are often non-compositional and culturally grounded. This study presents an idiom-focused diagnostic evaluation of English–Marathi neural machine translation using a corpus of 3,351 aligned sentence pairs and a smaller manually identified idiom-focused subset. Five translation systems are considered, including Bahdanau Seq2Seq, Luong Seq2Seq, a Transformer baseline, Helsinki-NLP opus-mt-en-mr, and a fine-tuned multilingual M2M-100 model. To complement surface-level metrics such as BLEU, METEOR, and chrF, the study uses Idiom Preservation Accuracy (IPA), a human-annotated measure that evaluates whether figurative meaning is preserved in Marathi translation. The available results suggest that fine-tuning improves idiom preservation in the diagnostic English–Marathi evaluation subset. However, because the idiom-focused subset is small and complete statistical outputs are not available for all reported metrics, the findings are interpreted as indicative rather than statistically conclusive. Overall, the study shows that IPA can serve as a useful complementary evaluation measure for assessing figurative meaning preservation beyond surface-level similarity.</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

Idiomatic english–marathi neural machine translation: a comparative study with an idiom preservation accuracy metric

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Jayanand A. Kamble,
  • Shivajirao M. Jadhav,
  • Vinod J. Kadam

摘要

Idiomatic expressions remain difficult for neural machine translation because their meanings are often non-compositional and culturally grounded. This study presents an idiom-focused diagnostic evaluation of English–Marathi neural machine translation using a corpus of 3,351 aligned sentence pairs and a smaller manually identified idiom-focused subset. Five translation systems are considered, including Bahdanau Seq2Seq, Luong Seq2Seq, a Transformer baseline, Helsinki-NLP opus-mt-en-mr, and a fine-tuned multilingual M2M-100 model. To complement surface-level metrics such as BLEU, METEOR, and chrF, the study uses Idiom Preservation Accuracy (IPA), a human-annotated measure that evaluates whether figurative meaning is preserved in Marathi translation. The available results suggest that fine-tuning improves idiom preservation in the diagnostic English–Marathi evaluation subset. However, because the idiom-focused subset is small and complete statistical outputs are not available for all reported metrics, the findings are interpreted as indicative rather than statistically conclusive. Overall, the study shows that IPA can serve as a useful complementary evaluation measure for assessing figurative meaning preservation beyond surface-level similarity.