Applications of Artificial Intelligence (AI) are diverse and can be used across the globe. For example, Machine Language (ML) Large Language Models (LLMs) are considered and studied for language translation in various contexts[1],[2],[3]. Use of automated facilities for machine translation typically requires some validation. However, such facilities can be useful tools, especially for large documents and translations to more than one language[4]. The NVIDIA NeMo™ Framework and Toolkit[5],[6] includes LLM support and translation models for several languages. The NVIDIA NeMo language translation models are based on the Transformer sequence-to-sequence architecture. This white paper provides an example of multiple-language translation of a short letter on a Dell PowerEdge 740 server[7] with an NVIDIA A100 GPU with 40 GB memory[8].

This white paper includes the example letter for translation from English to various languages using a parameterized recipient name. Depending on your environment and language support, some of the files might not be readable. The following figure shows an overview of the Finding Nemo English Translations and letter with the names that are substituted for each translation. A name that is associated with a specific target language replaces the ##NAME## parameter.

Figure 1.   Finding Nemo example

This white paper also describes an attempt to translate larger files and identify errors that are due to insufficient memory for such large datasets. For larger text files, GPU memory and configuration management are important. Such issues require additional preparation and configuration such as splitting data into smaller sizes and recombining the results, using GPUs with larger memory or using nondefault parameters. However, this information is beyond the scope of this white paper.

Notes:
Carefully evaluate and conform to all software license requirements. For more information about NVIDIA NeMo use and license scenarios, see Get Started With NVIDIA NeMo (https://www.nvidia.com/en-us/ai-data-science/products/nemo/get-started/) and https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/stable/starthere/intro.html#license. This white paper does not include licensing details for any component. 
 

Some software installation steps are shown for completeness. However, instructions change with software versions and are not the focus of this white paper. See relevant up-to-date sources for installation and configuration information.

[1]   Large Language Models in Machine Translation - ACL Anthology

[2] The Impact of Translation Shift and Method on Translation Accuracy Found at Bilingual History Textbook | Sipayung | Humaniora (ugm.ac.id)

[3] 2306-MTMA-LLMforMT (jhu.edu)

[4] Machine Translation Models — NVIDIA NeMo

[5] NeMo Framework for Generative AI - Get Started | NVIDIA Developer

[6] NeMo | NVIDIA NGC

[7] PowerEdge R740xd Rack Server | Dell USA

[8] NVIDIA A100 | NVIDIA