Machine translation (MT) often can save you money over exclusively human translation because machines translate more quickly. The key, of course, is to be sure the translation is accurate.
MT quality has improved significantly over the past few years. Beginning about five years ago, machines gained the artificial intelligence necessary to translate based on the context of the entire passage and not just phrase by phrase. Translation engines with AI capabilities can translate quite accurately in areas in which language is highly structured, such as the weather, sports, legal, and financial areas. The accuracy of your own MT, however, depends largely upon the translation engine you use.
Table of Contents:
2. Evaluating or Estimating Quality
3. Evaluation Methods
4. Estimating Quality
5. The Value of Multiple Engines
6. Human Touch
7. Andovar Offers Customized Translations and Localization Support
Types of Translation Engines
Translation engines generally come in two types. Both types use past translation data to translate the current passage.
Generic engines are those accessible to most people; for example, Amazon Translate, Google Translate, or Microsoft Translator. They are relatively inexpensive to set up and use and can convey the overall meaning of a passage, although they may not necessarily provide an accurate word-by-word translation. These engines are not customized to any specific industry or type of material but are designed to be used in various settings.
Custom engines work well if all your content is of a certain type. You provide data to help the custom engines refine their output. For example, if you are in the hospitality industry, you might "train" your engine to produce hotel listings a certain way each time. Custom engines can be quite accurate as long as the content remains the specific type the engine is trained for but are less accurate outside of the specific type. An engine trained to translate hotel listings would likely do a poor job of translating the weather, for example. These custom engines also are more expensive to set up and maintain than generic ones.
Engines also translate in different ways. For example, some use rules for certain languages to develop their translations. Others use AI or neural. Still, others use statistics to determine the most likely translation. Some use a combination of multiple methods.
Evaluating or Estimating Quality
You'll want to determine the quality of any search engine before you buy one to ensure you get your money's worth. You might choose to evaluate or estimate quality compared to that of human translators or to estimate it based on the source text. Either way will give you an idea of how well a particular engine will perform and help you determine the best engine for your business.
You evaluate quality based on one of two methods. The first is to have bilingual experts rate the quality of the MT output vs. human translators' output. This way of testing is expensive because human translators and human evaluators are required. Humans also are more likely to rate human translators more highly than machines because of inherent biases toward themselves.
A second way of evaluating MT quality is through relying on computer algorithms. The algorithm quickly scores the MT output against the reference translation. Algorithms differ. Among those used are:
- Bilingual Evaluation Understudy (BLEU)
- Recall-Oriented Understudy for Gisting Evaluation (ROUGE)
- Metric for Evaluation of Translation with Explicit ORdering (METEOR)
- LEPOR, which compares neural machine translations against that of humans and against that of other engines
Each algorithm uses a different approach to evaluating how similar the MT is to the human-translated reference passage. Quality evaluation can be a very effective way to compare engines, but it can also be slow and expensive. It also only evaluates quality at a specific point in time; since MT uses past translations as guides, they tend to improve substantially over time.
You can also estimate the quality of an MT. This is less expensive. Rather than measuring the actual similarity between a machine translation and a human one, this method analyzes the source text you wish to translate and, based on certain criteria, predicts the translation quality.
The Value of Multiple Engines
Since engines perform differently, many businesses will benefit from the option to use multiple engines. One type of engine might work better for one text, while another works better for a different text. Translation management software can switch from one engine to another depending on which is likely to provide the best translation. It can move between generic engines to custom engines as required.
No matter how good the MT is, you will still likely need some human touch to localize your message. Localization allows for the nuances involved in a particular culture and can differentiate between a text for Canada vs. a text for the United Kingdom, for example. MT also tends to be considerably less accurate when translating poetry and fiction. In these cases, however, a good MT can still save you money by translating the passage first, leaving difficult passages, and fine-tuning for humans.
Andovar Offers Customized Translations and Localization Support
Choose Andovar when you need translations that help you grow your global community of users and customers. From elearning to eCommerce, we work with numerous industries to provide marketing support in targeted regions. We also offer dubbing, voiceover, and translations for apps in our extensive portfolio of services.
Enter your email address to book an appointment and find out the return on investment possible with Andovar.
For more tips and content on global growth, please visit our blog.