Machine Translation and Human Translators

Machine Translation and Human Translators

Is Machine Translation good enough to replace human translators and interpreters?

This is a question about MT that comes up often. Conceived of in the mid-20th century, MT has improved exponentially in recent years, and there is a lot of investment by major corporations in its ongoing development. But currently MT is effective as a tool for linguists and laypeople, not a replacement for their minds. MT requires human involvement at many stages: initial programming, “training” it with data, creating strategies for its best use, and throughout the post-editing of MT-produced texts.

Basically, Machine Translation is computer software that decodes text from one language and delivers a re-coding of that text in the requested target language. It is different from Computer Assisted Translation (CAT) that human translators incorporate into their translation flow to quickly call up previously translated phrases, technical terms, and other stored language data. MT, on the other hand, takes on the role of the primary translator. It produces a finished text that is imperfect in varying degrees depending on many elements, including the language combination. Human translators then post-edit the text to create smooth final copy.

MT has some stumbling blocks that are hard to resolve without human assistance. Most notably, MT cannot handle:

1. Disambiguation (when a word or phrase has multiple potential meanings)
2. Non-standard speech, such as technical language or industry-specific jargon
3. Proper nouns
4. Translation between languages with different word orders

This is because, as excellent and complex as MT has become, its processes are not actually those of thought. MT and other forms of AI may provide super-fast, relevant “answers,” but they do not think. In other words, they can’t make nuanced choices because they can’t surmise context. They only compare and match patterns. Therefore, MT is not terribly helpful for creative texts like marketing and literary translation.

Some examples of good uses for MT are translations of customer support communication, assistance content, and user documentation. For the European Commission, one of the largest institutional MT users, translation assistance for the constant flow of often-repetitive EU texts is invaluable. MT mobile app use is high in organizations with field operations, such as militaries. The explosive level of social networking online is another source of demand for MT. In all of these instances, approximate translations are usually okay.

But when they’re not, MT can still provide value by shortening turnaround time, improving the accuracy of repetitive portions of your text, and significantly lowering the total cost of your translation project. If your Language Service Provider offers translation aided by MT, it will be one of the three major types of MT, or some combination:

1. Rule-based (RBMT) – Using linguistic rules and bilingual dictionaries, RBMT replaces source language words and phrases with best-matches from the target language.

2. Statistical (SMT) – Algorithms select the most likely target language text based on analysis of the “training” data, which works best for languages that use the same word order.

3. Neural (NMT) – A type of MT only in common use since 2016, NMT operates like the extensive neural networks of the human mind, executing a form of AI “deep learning” that finds statistically likely target text through a more complex process than SMT (how it does this is not completely understood).

All of these systems require extensive investment by an LSP – investment in the software and hardware systems themselves and also the training of the personnel who will use them. MT systems are more effective the more you use them, so they can be part of a long-term plan to improve your translations while decreasing their cost. In fact you can contribute to the efficiency and quality of your projects’ execution by providing an organized glossary and original texts containing simple, clean, repetitive prose.

In order to determine whether MT would be a good choice for your particular translations, LSPs use established methods to examine and score them. And if you are doing anything more than the most rudimentary translation of data, consider requesting that human translators be involved from the beginning through a final, thorough post-editing process.