Will innovations in Neural machine translation technology herald the end for professional translators, or help fill in the gaps?
In almost any industry these days, technology tends to come up in table talk as the big bad wolf that’s coming to eat up your job. While this may be the cause for alarm for several professions (accountants, salespersons, and office clerks, beware!), most still require a human touch that no AI is yet able to replicate.
Recent inroads in neural machine translation have brought AI a significant step toward parity with human translation, but will this impact the translation industry’s bottom line? To learn more about the machine translation trends in this industry, keep on reading!
The Current Landscape of Machine Translation
Before we get started, let’s first define what it is? At Tomedes, we describe Machine translation (MT) as any translation process that solely relies on a computer program without human involvement in the translation.
Since the first successful translation from Russian to English in 1954, MT has come a long way. A good example of how far we have come is that last 2020, Facebook introduced a new AI model for neural machine translation (NMT) and made it available open-source. Dubbed M2M-100, this NMT model can translate between any pair from among 100 languages without using English as an intermediary.
Facebook’s model uses Convoluted Neural Network (CNN) architecture, allowing for more flexible non-linear computation of linguistic data than the Recurrent Neural Networks (RNN) that are the industry standard, making it the first truly multilingual translation system. In terms of performance, it scores a full 10 points higher on the BLEU metric than English-centric models.
While it isn’t currently in use, it’s poised to be a big leap in the machine translation trends since Google’s own patented GNMT system was introduced in 2016. Google, of course, is not one to be upstaged, having upgraded from its original RNN model to a hybrid built primarily on better-performing transformer architecture.
The Machine Translation Trends: a Bane or Boon for Professional Translation?
There’s a thicket of specialized tech talk around the finer details of these developments. Still, the main takeaway is clear: machine-based translation is becoming more accurate, more responsive, and more natural-sounding than ever. But what does this mean for the translation industry?
It’s not as grim as one might think.
The language service industry doesn’t need to have an adversarial relationship with technology. After all, the commercial availability of computers paved the way for machine-assisted translation software, which enabled professional translators to work more efficiently through translation memory tools, terminology banks, and electronic dictionaries, among many other devices.
Why in fact, way before 1954, the founding narrative of translation tech goes back to the 9th century with Arabic cryptographer Al-Kindi, whose frequency analysis method formed the base of modern MT!
But attempts at actual automated translation have never reached the required level of sophistication to replace human labor. Word-based and phrase-based attempts remain inaccurate, grammatically unsound, and for the most part, amusingly incomprehensible.
The idea that machines could even get close to parity with human translations was unthinkable until 2016 when Google replaced its decade-old predictive algorithm with a patented neural machine system that would set the industry standard for the decade to come.
Current Issues in Machine Translation
The switch to Neural machine translation models has exponentially improved the quality of MT systems. But despite their sophistication, there are still several drawbacks to MT that ensure the continued need for thorough human oversight in the professional setting.
- Performance on resource-poor languages: Because most NMT models have been trained mostly on data in English or use English as an intermediary to bridge other language pairs, MT tends to perform better when translating between English and another language. This impacts the quality of translations in proportion to the rarity of the language. Facebook’s model currently best represents the effort to solve these particular issues.
- Lack of creative and cultural nuance: NMTs still treat language as a set of computational vectors and can only infer context and meaning from the text as given, no matter how natural the output may appear. This means that the more complex and idiosyncratic uses of language, as in literature, or cultural connotations that affect the use of language, still tend to be lost in translation.
- Professional standards: The language services industry is highly competitive, and the level of linguistic sophistication clients demand often exceeds the capabilities of even the most state-of-the-art MT tech currently available. In addition, translation projects across many fields such as law, medicine, and finance cannot be scaled beyond the exacting scrutiny of a professional, as even one minor error could lead to disastrous results.
Due to the mentioned problems with the current NMT models when it comes to the translation process, machine learning engineers and linguists are vital in developing and studying these technologies to make them more accurate.
Machine Translation Post-Editing
But this isn’t to say that MT has no use in the professional setting. While the human element remains of paramount necessity, MT has become a viable tool to help translators work more efficiently in the past three or four years. The key here is Machine Translation Post-Editing (MTPE).
MTPE is the process of running a text to be translated through an MT system, after which a professional translator combs through the translated text to ensure the translation is correct and comprehensible.
MTPE is often categorized into two grades, light post-editing, and full post-editing. In light of post-editing, a translator will correct mistranslations and ensure general readability to convey the gist of a document and is the faster and more cost-effective option. Full post-editing ensures that the translated text is of the highest possible quality, addresses stylistic errors, adapts proper tone and phrasing, and makes appropriate adjustments for cultural fit and creative use of language.
At Tomedes, we include a third grade: specialist post-editing, in which the translator assigned for full post-editing is also an expert in the field relevant to the document to be translated. This ensures that specialist terminology, linguistic conventions, and complex concepts are dealt with appropriately during the translation process.
The Future of Translation with Machine Learning
Ironically, these developments in machine translation may be helping the human element of translation shine through. Because it can help cut through the more rote and repetitive aspects of translation, translators can put more of their energy into delivering a polished, carefully-made product. This means better services in specialized fields that demand precision of language and in more creative applications, such as marketing.
Better machine translation also means a better experience for users in everyday settings. Things like translating an article for casual reading, items on a restaurant menu, or learning how to say particular sentences in another language—these are things that people wouldn’t normally hire a professional translator to do anyway. Because of this, machine learning engineers and linguists continue to develop new algorithms and frameworks for us to better communicate with anyone in the world.
As such, the rise of AI in translation is not something that professional translators need to fear, even well into the future. It is exciting because as translation technology evolves, so does the industry and everything it can offer.
Author Bio – Ofer Tirosh is the CEO of Tomedes, a translation company with expertise in machine translation post-editing solutions. From the outset, Tomedes has always embraced technology that could help translators provide top-of-the-line service with consistency and sustainability in mind, with a decade of innovation and service to over 95,000 clients in 120 languages and 950+ language pairs.
More Stories
Apple to expand end-to-end encryption to more iCloud data categories
Back to Basics With Email and RSS
VentureBeat Lab to launch in-depth AI article series with insights from Microsoft and Nvidia