Maximizing Your ROI: Strategies for Launching NMT | Argo Translation

In a recent blog post, we discussed the misconceptions about using Neural Machine Translation (NMT) and generative AI to improve the translation process. This blog post outlines how to launch a proper strategy using NMT.

Understanding the three types of NMT engines, standard, trained, and custom, is crucial to understanding your final strategy options.

Definition of a standard NMT engine

A "standard" NMT engine would be a general-purpose translation system that is not specialized for a particular domain or type of text. It's built to perform reasonably well across a wide range of texts. In contrast, a custom NMT engine might be trained on specialized corpora, like legal or medical documents, to perform better on those specific types of content.

Definition of a trained NMT engine

A trained NMT engine is a neural network-based system trained on a substantial dataset comprising pairs of sentences in two languages (source and target). The engine learns to predict the sequence of words in the target language corresponding to a given sequence in the source language through the training process. It achieves this understanding by optimizing its internal parameters to minimize the difference between its predictions and the actual translations in the training dataset. This is the best option if you have a large set of human-verified translations (greater than ten thousand segments). This is the preferred and most accurate approach of the three models.

Definition of a custom NMT engine

A custom NMT engine is a specialized, tailored version of an NMT system designed, trained, and optimized to meet specific translation needs or requirements. Unlike standard NMT engines that aim for broad applicability across various languages and domains, custom NMT engines focus on delivering high-quality translations for a particular language pair, domain, or even style based on the unique dataset trained on and the specific configurations applied during their development. This is the best option if you have a limited amount of human-verified translation, but you do have glossaries and less than ten thousand segments of human-verified translation.

Your approach depends on the type and amount of human-verified translation you have completed or plan to complete before starting the project. Your language service provider can help you determine the most effective path forward.

As AI continues to play a bigger role in translation, understanding how it fits into your specific workflow is key. For a deeper look at how AI translation can streamline your projects and improve efficiency, explore our AI translation eBook for practical insights and strategies.

Next, let us understand why you and your organization might consider an NMT+PE (Neural Machine Translation + Human Post Editing) strategy. Increasing demands on organizations to provide more localized content faster have created the need for alternatives to typical human translation. Adding NMT can provide a healthy set of options where you can match the desired level of quality to the proper service and cost level.

*Using a free NMT engine typically does not guarantee data privacy and security. However, using the paid API version typically ensures zero data trace privacy.  

Here are a few benefits of pursuing an NMT+PE strategy as part of the continuum of translation services you purchase.

1. Higher Translation Quality: NMT models are trained on large datasets and can capture complex linguistic patterns, resulting in translations that are often more accurate and fluent than rule-based or statistical approaches. For many content types, the quality is comparable to human translation if an adequately trained NMT engine is paired with a good editor for the review pass.
2. Cost-effectiveness: Although a trained NMT model requires some initial setup costs to train, once the model is trained, the price per translation can be lower than traditional translation services, especially for large volumes of text. If you do not have enough historical translation, you may need to invest in having at least 10,000 segments of human-verified translation. This volume is generally the minimum translation recommended for adequately training your NMT engine.
3. Faster Turnaround Time: NMT systems can translate large volumes of text quickly, which can be particularly beneficial for companies that require rapid translations to keep up with the pace of their business operations.
4. Customization: Companies can fine-tune NMT models to their specific domain or industry, making translations more tailored to their needs and terminology.
5. Scalability: NMT systems can quickly scale to handle increasing volumes of translation tasks without significant additional resources, making them suitable for companies with growing translation needs.
6. Consistency: NMT models provide consistent translations for similar text inputs, ensuring uniformity across translated documents, which can be crucial for maintaining brand consistency and clarity of communication.
7. Support for Multiple Languages: NMT systems can support translation between various language pairs, making them suitable for companies operating in diverse linguistic environments.
8. Privacy and Security: Since NMT systems can be deployed locally or on private servers, companies can have greater control over their translation data. This control addresses potential privacy and security concerns associated with sending sensitive information to third-party translation services.
9. Integration with Workflows: NMT systems can be integrated into existing company workflows and applications, allowing seamless translation processes without manual intervention.
10. Adaptability: NMT models can continuously learn and improve with additional data, allowing companies to adapt to evolving language usage and terminology within their industries.

The impact of a properly trained NMT engine

The best way to assess the impact of this type of initiative on your organization is to review an example. For this example, we will look at a manufacturing client we have worked with for 29 years. Their translation memory contains over 300,000 human-verified segments per language.

We built a trained NMT engine using Azure and the corpora of human-verified translation. The results showed a significant improvement compared to the untrained NMT engine. We have two measures we can use to measure the engine's efficacy.

The first measure is the Bilingual Evaluation Understudy (BLEU) Score, a metric used to evaluate the quality of machine-translated text compared to human translations. It measures the similarity between the machine and human translations by calculating the precision of matched words or phrases, considering their order and appropriate weighting, to produce a score ranging from 0 to 1, where 1 indicates a perfect match with the reference translation.

Before training, the NMT engine baseline BLEU score was 0.58; after training, the score improved to 0.89.

The second measure is based on the number of changes an editor makes to the machine-translated content. We used a representative document containing 50 segments. The score for the translation from the untrained engine comes in at 4.6 out of 10. The editor modified 27 out of 50 segments. The score for the translation from the trained engine comes in at 9.0 out of 10. The editor modified only 5 out of 50 segments. The errors were simple capitalization issues. The increase in quality is significant. This NMT engine is very effective and ready for deployment.

Before training

After training

In closing, most translation buyers can find a suitable use case for NMT+PE as part of their translation strategy. Implementing the plan can bring many benefits, including decreased timelines and costs.