The Intento MT Studio for in-depth, data-driven MT evaluations gives you the opportunity to select and improve the machine translation models that will have the greatest impact on your specific project context. You can evaluate your own models, or choose from pre-trained stock model offerings from the 40+ most popular MT providers available on the Intento platform.
An Intento MT Studio subscription takes your previously translated content (along with any translation memories and glossaries), and ultimately allows you to save time and money on post-editing effort and manage quality for real-time cases such as support chats and internal communications. This all happens in one simple interface where you can handle your training and evaluations processes for as many providers as necessary.
We’ve introduced three new features to the Intento MT Studio, all aimed at solving some of the most common requests from those working throughout the MT landscape. Data cleaning and customization analysis have both been enhanced, giving you a seamless post-editing process — no matter how many resources you have at your disposal. At the same time, we’ve integrated the entire workflow into one seamless interface, taking away the need to import or export data throughout the project.
Keep reading for an overview of everything new in the Studio:
- Data Cleaning update
- COMET-QE metric, with or without human reference
- More efficient evaluation of trained models
– and here’s to higher acceptance rates with less effort.
• • •
Data Cleaning
At Intento, customers are asking how to clean data properly in order to get the best options for model customization. Unwanted or redundant information in the translation memory negatively affects the quality of the resulting model. This often makes data cleaning a necessary part of the process for custom model training. Every issue that you see in your data-set will come up again as problems in the custom model. Eliminating these issues is the only way to experience near-perfect results. Seeing that there was no easy-to-use tool on the market, our experts built MT Studio Data Cleaning. Simplifying the most difficult and time-consuming aspects of the process, Data Cleaning is the first all-in-one tool for quick and easy data prep in order to optimize custom model outcomes.
The focus is on several areas; removing duplicates, emojis, HTML makeup, and running the numbers in the source text compared to the target. All of these aspects of model training are very difficult to handle on your own, which is why MT Studio fits them all into one tool. This way, you can get better results with less effort. Select the parameters by which Data Cleaning will modify your datasets. If you’re new to custom model training, leave all the filters on and you’ll have nothing to worry about. However, feel free to modify which filters are applied in the Studio for specific use-cases. In the next release, you will not only receive the polished data, but also a catalog of removed segments with explanations as to why each segment would hinder the training process.
• • •
New metric: COMET-QE — Check MT quality without human translation reference
When you want to estimate how well a model is translating content, most of the metrics are used to compare the MT output to the ‘golden’ reference of human translation. However, in many cases, the customer may not have this ‘golden’ reference to leverage. Maybe there is a new language pair that the customer has never used before, or an LSP is handling new clients with unique requests. Even for real-time cases such as support tickets or online meetings, you won’t have the human reference, but you still will want to know what to expect from the MT. That’s why Intento MT Studio employs scores that predict the result of different models applied to any use case.
When using MT Studio, you will now have access to the COMET-QE metric to compare output from different providers with or without a human reference. You can find the best provider for your project no matter how much data is at your disposal. With the new metric option, MT Studio now has 6 metrics to assess the quality of your models:
• • •
Evaluate trained models more efficiently
MT Studio is an entire toolkit to enhance your MT experience, leveraging everything from training to scoring to cut down post-editing efforts. You can now handle the entire workflow from beginning to end in the same place. This means cleaning data, training models, analyzing models — all without a single interface without importing or exporting your data. You can choose to export data at any point, but a continuous workflow is available for any Studio project. You won’t even need to leave the MT Studio for translations; we’ve added that as well.
