Error Analysis of Machine Translation for Malayalam Fiction

Machine translation (MT) has transformed translation studies and linguistics, significantly improving cross-cultural communication and linguistic analysis. The aim of this study is to evaluate and compare the accuracy of Google Translate, AI4Bharat’s IndicTrans2 and Bing in handling Malayalam compound nouns, with a particular focus on Named Entity Deviation Errors. The study seeks to identify specific challenges in translating Malayalam noun formations and case markers and to understand their impact on translation quality. Utilizing a mixed-methods approach, the research involves quantitative and qualitative analyses of three corpora built from a selected fiction text in Malayalam and its human English translation. Our findings revealed significant issues in translation accuracy, and some common errors were identified including improper translations of proper nouns, mistranslations of compound nouns and transliteration issues. Automated metrics used to analyse errors in each MT model revealed that literary-adapted machine translation models produced richer output and showed improved performance compared to general domain models. The study accentuates the necessity of robust linguistic models and larger, high-quality parallel corpora to enhance MT accuracy for low-resource languages like Malayalam. The study suggests a hybrid approach that develops MT to achieve greater precision and reliability in translation practices, ensuring nuanced and contextually accurate translations.