<![CDATA[Multilingual language models (MLMs) are widely used for cross-lingual tasks, yet their ability to achieve consistent semantic alignment and transfer to low-resource languages remains limited. This work examines cross-lingual semantic alignment and transfer learning through a comparative evaluation of MLMs at both the word and sentence levels. We analyze general-purpose models such as BLOOM and task-specialized models including LaBSE and XLM-R across English, French, Hindi, and Kannada. Word-level experiments show that LaBSE achieves substantially higher cosine similarity scores of above 0.80 across languages. In sentence-level natural language inference, XLM-R outperforms other models, achieving an F1 score of 68.62% on Kannada and 74.81% on French. These results indicate that model specialization and training objectives play a crucial role in cross-lingual performance, particularly for low-resource languages, and should be carefully considered when deploying multilingual natural language processing (NLP) systems.]]>
Copyrights © 2026
