1. “You mean Google Translate, right? ”
The first thing that pops to mind when people hear machine translation, is Google Translate. It's the most well-known, publicly available tool, and accessible to all. Therefore, some people tend to use Google Translate as a synonym for machine translation, with funny translation hiccups coming to mind. It's true that Google Translate is based on neural machine translation (NMT), deep learning, and neural networks as well. However, the different ways in which AI systems are used result in massive differences in quality. The online tool, for example, uses generic data, which it collects from all over the web. This in turn produces generic translations, that might help in understanding the essence of a text. However, when training NMT using specific data, the technology starts producing much better results. Together with quality checks by human translators, the tech can also be used for the professional use case.
2. “Machine translations will replace human translators”
No! People can do better and more productive work with the support of innovative technology. This has always been the case, and translations of texts are no exception. We will always need humans to check the results, because even the best machine translations will never be perfect. The tech does help in increasing the volume of texts that can be translated. “There are around 300,000 translators on the planet, and this figure has remained constant over the past years," explains Jay Marciano, our Director of MT Strategy & Outreach and Board Member of the AMTA (Association for Machine Translation in the Americas). "On the other hand, the quantity of content is simply skyrocketing. Every year we produce 40 percent more data than we did the year before. So the amount of translatable content is increasing much faster than the volume that professional translators can handle. Tech can help boost translation volumes, to meet the steady rise of demand for translations.
3. “MT produces poor results”
MT comes in a variety of flavors. Whenever we speak of MT, we actually mean neural machine translation. This is the state-of-the-art method currently available in the field of machine translation. The “legacy model”, statistical machine translation, was the gold standard in the industry for quite some time and it works like this: computer systems read words and sentences in the source and target languages and are given a set of programmed language rules. They use this input data to produce those translations which are the most likely option for corresponding words and sentences.
Similar to statistical machine translation, NMT also leverages historic translation data, stored in translation memories, for example. However, NMT teaches itself the relationships between the source and target language, using deep learning methods, doing away with pre-programmed rules. On this foundation, the systems establish language models made of neural networks, similar to those found in the human brain. As such, the computer independently identifies correlations between the target and source languages, virtually teaching itself the rules. This comes with one major advantage: The systems can be finetuned to a high extent.
NMT produces even better results when language models are exclusively trained using only relevant data (see Point 1). When translators revise the machine translation output data is created in the process. This interaction data is in turn used to teach the machine to do its job even better. This massively improves the translation quality and produces results much quicker over time.
4. “Human translators will always outperform MT”
It's not a matter of (wo)man vs. machine. Both humans and machines bring different skills to the table when it comes to translating. Humans are a diverse group, with each human having a distinct style of translating, seeing as how every word in the source language can be translated in a variety of ways in the target language. MT technology can ensure great consistency – uniform translations – also when working on many different documents. Human translators are capable of detecting fine nuances in language, and check whether these nuances are reproduced correctly in the target language. The best results are produced whenever people team up with machines, complementing each other's skill sets.