Machine Translation

MT is an area of the localization and translation industry that many companies have yet to crack, but so far, we’re on the verge of looking at ways in which it can assist or even help us fully in our translation efforts. Machine translation is quite self-explanatory, it entails the process of translating words from one language to another via the sole means of a computer system or artificial intelligence. Some may see this concept and be instantly turned away by the potential ideas of errors, misunderstandings and miscommunications created due to the translation process moving entirely through that of a computer. However, nowadays the idea of MT is more than plausible, especially when we consider that of human post-editing (MTPE), whereby professionally trained linguists or translators, analyse the MT data output, fixing any inaccuracies by correcting and resolving any linguistic or semantic errors.

In the following blog we’ll look at a few things: further depth of what machine translation really is, its capabilities, its evolution, as well as who may and may not want to be using it for their documents. Ultimately, the general consensus is that machine translation is extremely viable for short-form writing; a few pages of a script or presentation will translate near-perfectly, be that through an online service or a more elaborate process. From there, in regard to longer or far more complex pieces such as technical or medical documentation, we then ask the question of, how do we go beyond? What companies can offer more? Ultimately, what are the steps they take in order to achieve that higher level of machine translation.

Uses and Limitations

Machine translation is a great blanket process for getting the heavy lifting of translation completed and in a short period of time, without the requirement of extra human input. This is where our first use comes into play in regards to time-management. Sometimes in the industry, be it a local translation service, or even a global localization company such as Janus Worldwide Inc, providers require quick turnarounds for a general translation of a document, briefing or other text-based project. When we look to this context we see the ideal scenario where machine translation can be used: short-piece documents or work that only the “gist” of the piece needs to be conveyed is where MT flourishes. However, beyond that, perhaps in the realm of technical or medical documentation, we see our first limitation, being that of specifics.

In the context of machine editing, the barebones, no human input required process is often referred to as “good enough” quality. This implies that the meaning and intention of the text can be conveyed, the semantics, basic rules of language, and some parts of the tone will make it through in translation. However, for those instances where nuance, detail and specifics are required, we must enlist the help of humans in the areas before and after machine editing; either that, or we look to more complex means of MT such as neural or statistical. Neural machine translation and statistical machine translation are the two cutting edge methods of the more complex side of the translation industry, for a few reasons. The first of which being that they rely on numbers, and mathematical assignment as opposed to merely relying on the words themselves. Of course, the words do hold strength, and it’s strange to think that translation can be handled through equation, but to simplify the process: with both neural and statistical machine translation, each word is assigned a number, and the process moves on from there. The difference, and reason why NMT is often favoured over SMT, is that the former has chosen the route of giving similar words similar numbers, helping the machine understand words in a more digestible manner. As a result of these more complex means of machine translation, we reach a new definition often referred to as “near or matching human translation”. Here, we find that all the factors we mentioned previously, such as nuance, grammar and semantics are all fine, as well as knowing that the document’s syntax, formatting and black or white-listed words have not crept their way in.

Pitfalls and Solutions

One pitfall of course refers to the inaccuracies that can occur when using machine translation, and the second would be the time and human effort it would take to correct or prevent these inaccuracies. As aforementioned, when looking at the simpler, less costly side of MT, we are often faced with these inaccuracies, and fill in the issue with human input. However, the fix for any issues and kinks made by the machine during the process are easy to dispose of with a few good tricks. The first thing we need for our list of solutions to work is a competent linguist technician who can tamper with “the before” and “the after” part of the process. Though it’s important to remember that this is only for the lower-end of the MT spectrum, the higher quality services used, and debatably more money spent, the better result will be achieved.

Pre-Editing and Post-Editing

The work done on the machine before is what we call “pre-editing” in which the parameters and settings of the machine are modified to meet the requirements of the client. Perhaps there is a list of words that we do not want to be featured in the translation, or a word that we do want, that will help convey the company’s message. This is the simpler side of the process, there are other, more complicated additions relevant to the pre-editing phase such as the use of translation memories, data banks, and pre-existing assets owned by the creators of the machine. In comparison to post-editing, the pre-editing stage can require more human effort and time, as it involves a lot of setup and will be different for every client. Whereas the post-editing stage’s importance lies within the simpler task of proof-reading, making sure that the translation was successful, none of those words crept through the filter, and that the overall message remains intact.

Due to the nature of machine editing, another pitfall that shows up quite a lot is payment for both pre and post-editing processes. More often than not, pre-editing is very bespoke to the company, and therefore requires care in setting up, and dependent on the document type, post-editing can take an extended period of time for correction and checking. However, the solution is fairly simple in that these areas of work can be compared to other business models, and compensation for the process can be applied similarly, for example, post-editing is a lot like proof-reading, and can be charged as such on an hourly basis.

Janus Worldwide Inc’s approach to machine translation is unlike any other, as we are fully committed to the idea of customization and client personalization of the project at hand. Offering creation and deployment of machine translation that can focus on any number of factors such as: writing styles, grammatical requirements, corporate-specific terminology and diction, special rules and formatting styles such as numbering or dates, as well as any other small details a customer may be interested in. All in all, saving the client time and effort that can be put into other means, whilst also keeping a competitive and fair price for the development and deployment of the process.

How do Janus pull off such a high-quality job? Well, Janus uses previously created company assets that can be reworked and put back into other projects, such as machine translation. Utilizing a range of different language combinations and data, such as their pre-existing translation memories and lingual databases to provide efficiency in their delivery. However, having the assets at hand simply isn’t enough, to fully utilize said data, Janus puts their world-class linguists at work to pre-train the engine that will be translating the clients’ writing with methods such as corrective feedback. From here we can relate the next step back to our previous paragraph on the different needs and uses of machine translation, some clients may only need a basic translation, so that the document can be understood by an audience in its general sense. Others may require MTPE, which Janus can also help with if the client so chooses.

In Conclusion

Overall, we’ve looked at the ins-and-outs of machine translation, post-editing and even pre-editing, as well as the differing layers that fit in and around the translation process, and more in-depth services such as NMT (neural machine translation) and SMT (statistical machine translation). We’ve gone over the uses, limitations and solutions for machine translation, as well as the appropriation of price in regard to other business models. Lastly, we’ve spoken about Janus Worldwide’s take on MTPE, and what we are doing to ensure we capitalise on this ever-evolving market in a way that uses our expertise to help clients with high-quality translation at a fair price.