What are the most common misconceptions about machine translation?

Misconceptions regarding machine translation (MT) include its ability to completely comprehend context and cultural nuances, which often leads to an overestimation of its proficiency. Many people believe that machine learning works flawlessly across all content kinds, neglecting the fact that its performance differs depending on language pairs and subject complexity. Furthermore, the illusion persists that MT does not require human intervention, despite the fact that post-editing is critical for correctness. Understanding these myths will help you set reasonable expectations, optimize your MT use, and recognize its limitations.

A human translation will absolutely take context, voice, brand integrity into account. MT has it's pace however it has to be the right machine created for the right application, for the right content. After all it's a machine. Generic machines might help some understanding. However they will confuse a professional translation and dilute your brand and message.

For companies considering the move towards using MT, be very aware of the perils of unsupervised (i.e. non post-editted) machine translation. Since post-editting of MT can be soul-destroying for professional translators, think carefully about involving them in any strategic decision, and not relegating them to a post-editor. Adopt an expert in the lead approach (i.e. machine in the loop) rather than relegating the expert to being the human in the loop. Last but not least, remember that MT is not a panacea - there are many instances, particularly in creative tasks, where MT is not fit for purpose. Don't risk your company's reputation by being over-reliant on MT.

Machine translation is always faster than human translation. While machine translation can be faster than human translation in some cases, it's not always the case. Machine translation tools can struggle with certain types of content, which can slow down the process. For example, if you're translating legal or technical documents, you may need to spend extra time reviewing and editing the machine translation to ensure accuracy.

Whereas I agree with the overall idea of this being a myth, my main red flag is for how long. MT can certainly struggle with idioms and metaphors, or offering quality assurance. So far. With the arrival of AI we have tools capable of producing good translation results, as well as insightful feedback. To a certain level it is already possible to translate many of those essential nuances to an extent that it was impossible to think just a year or two ago. Therefore, I understand "can replace" is a myth in its "present tense" meaning. But I have my doubts that "can replace" as in "having the ability of" is a myth or not.

Much like with human-generated translation, there needs to be a quality assurance process in place. MT is an efficient place to start, but humans are needed for QA and to ensure culturally relevant nuances are properly conveyed.

MT will frequently retain a word order in the target language that is very close to the source text. This may affect the flow of texts. Particular care may need to be taken when translating between inflected and non-inflected languages as well as languages with non-formal and formal registers, or with different frequencies of usage of the passive voice or subjunctive mood, or where there is no distinct tense mapping between the source and target languages.

Machine translation is perfect and can replace human translators. While machine translation has come a long way in recent years, it's still not perfect. Machine translation tools can struggle with idiomatic expressions, cultural nuances, and complex sentence structures. For example, consider the following sentence in Spanish: "Tengo un resfriado y me duele el estómago." A machine translation might produce something like "I have a cold and my stomach hurts," which is technically accurate but doesn't convey the same level of detail and nuance as a human translation.

For some odd reason, agencies employing MT always expect their suppliers to slash their rates. Their reasoning is that the machine has done the heavy lifting, allowing the linguist to blaze through the copy. What's iffy about that common misconception is that the mental drain of MTPE is much higher than that of regular translation. You still have to read the source and compare it to the MT's output, but you'll have to be twice as fast to earn the same income. After a couple hours of patching up shoddy MT output, your brain is fried – and tomorrow you'll get to do it all again, for pennies! As wonderfully impressive a technological feat it may be, its commercial application reeks of a sweatshop mentality: cheap, cheaper, cheapest.

Prescriptive and standardised texts may seem to be cheap to translate using MT. However, where terminology is not applied consistently, it means that nearly every sentence will require post-editing. Unless preferred terminology can be embedded, there will be particular deficiencies that increase the amount of segments needing to be revised. Particular problems also arise in more inflected languages, where gender/case agreements need revising, or in languages where informal/formal registers are mixed in the output.

Editing entails: ● Considering the text as a whole, not just sentence by sentence ● Correcting translation errors, addressing false friends, and rectifying incorrect context ● Checking for missing content in the machine's output ● Checking and maintaining consistency of terminology ● Correcting spelling errors ● Ensuring proper capitalization and adherence to style rules ● Verifying names, brands, and DNT terms ● Researching and standardizing terminology ● Dealing with cultural nuances ● Adapting speech and tone to the target reader. It therefore takes as long or longer than translation and cannot be paid for at a lower price.

Machine translation is free. While there are some free machine translation tools available online, they often come with limitations. For example, you may be limited in the number of words or pages you can translate per day. Additionally, if you want to use a more advanced machine translation tool, you may need to pay for a subscription or license.

Particularly in Low Resource Languages (LRLs), machine translation is very vulnerable to lexical gaps and semantic bleaching, in particular when translation from language A to language B is done through a pivot language (frequently English). Imagine translating a text from Aleutian to Saami through English about prevailing conditions in winter. The source and target languages both have many words for snow (e.g. differentiating between texture, density etc) but there is a risk of much of the meaning being lost in translation through the pivot language. This can lead to considerable translation loss, stripping out much of the cultural richness of the source language in the translation, and therefore much of the author's intention.

Machine translation (MT) is particularly deficient when it comes to European Portuguese and Brazilian Portuguese. Existing machine translators are generalized and not adapted to specific fields of knowledge. For example, it is not sufficient for a machine translator to have been trained on texts from the European Union, which is autonomous from the law of its member states, and completely apart from the law of countries belonging to different legal families. Machine translation is incapable of performing comparative law operations, a task reserved for translators specializing in legal translation, including lawyer-linguists.

Machine translation is only useful for low-stakes content. While it's true that machine translation can struggle with certain types of content, it can still be useful for a wide range of content types. For example, if you're translating social media posts or product descriptions, machine translation can be a quick and easy way to get the job done. Just be sure to review and edit the machine translation to ensure accuracy before publishing.

The appearance of MT being a black box is only the case for those who don't understand how an engine/model is trained. Ultimately models and engines use probability of occurrence and paired language data from source and target languages (if need be via a pivot language) to "translate". Different engines and models in the same language pair may yield different results due to differences in their training data, e.g. their domain specificity or register (e.g. formality) - to provide an example of formality, two engines might provide vastly different results in a language pair with formal and informal pronouns of address, although this can throw up interesting results if formal/informal data is mixed up together.

No one talks about how boring MTPE is. Translators do what they do out of a love for languages. They're craftsmen. The only thing MT can do is come up with literal options. By extension, the tech is useless to any translator worth their salt. MTPE is like the Henry Ford of the translation industry. We'll all end up producing the same dull and cringeworthy boilerplate copy – but at least the top-hat fat cats can proudly proclaim they're growing their business, fuelling innovation and moving the trade forward! As such, the potential mass adoption of MTPE will result in an exodus of knowledgeable translators, leaving us with the tyranny of empty-shell translation agencies subsisting on poor-quality, neither-here-nor-there 'vendors'.

It's important to remember that machine translation is a tool, not a replacement for human translators. By understanding the strengths and limitations of machine translation, you can use it effectively as part of your overall translation strategy.

MTPE is often described as boring and numbing by many in the linguist community. These linguists are qualified translators with diverse backgrounds and expertise. They are experts in their respective fields, using various digital tools, including machine translation. They are not the new Luddites but linguists who control their workflow and advocate for a 'machine-in-the-loop' (#mitl) approach rather than 'human-in-the-loop.' MTPE is known to produce low-quality and inconsistent texts with a limited vocabulary. Informed clients are aware that MTPE is rarely a suitable solution for their needs. Additionally, the impact of MTPE on translators' mental health and well-being has yet to be thoroughly studied.

In my experience, MT has proven itself helpful. However, it has its limits and reviews and post-edits are a must if you want to give a soul to the text. In the case of my mother tongue, it is not included in any MT, thus we mostly rely on the traditional way.

Machine interpreting is frequently touted, but is still a myth. The process touted as "machine interpreting" actually involves speech-to-text conversion and then machine translation, followed a text-to-speech transformation of the output. Its quality naturally depends on the initial speech-to-text conversion being able to eliminate "false starts", understand regional accents and pronunciation and different registers of speech, followed by then translating it accurately. The skill in human interpreting lies in identifying intonation, gestures, emotions, and facial expressions, as well as the ability to correct mistakes on the go as they become apparent, as greater context is obtained, all while not imparting bias in choice of words.