How does AI translation actually work and can it be trusted?
AI translation offers both speed and lower costs – but how does it actually work and how is the quality of an AI-assisted translation?
AI technology opens up a world of possibilities. It can save time and money – for both businesses and individuals. But how good is AI at translating? How does AI translation work?
Letting a machine translate is nothing new. For many years translators have had access to so-called machine translation to assist them. In fact, professional translation tools are commonly used in the translation industry.
In this article we’ll provide a brief overview of the following:
- How a translation tool works.
- How traditional machine translation works.
- How AI translation works.
- How to spot the pitfalls with AI translation.
In another article we’ll delve a little deeper into the pros and cons of AI translation.
What is a translation tool?
There are many translation tools (so-called computer assisted translation or CAT tools) on the market. Most of them have the same basic functions designed to facilitate translation work. We’ll briefly describe two of these functions: term base and translation memory.
Before starting a translation assignment the translator can build a term base by inputting words and phrases in both the original language (source) and the language the text will be translated into (target). This ensures, for example, that every instance of an industry-specific word, term or phrase in the text is translated correctly and consistently. Once the translator gets started the tool highlights the words and phrases in the source text that are included the term base and shows the correct translation.
The CAT tool also has a translation memory. Everything the translator has already translated is saved in the tool’s memory. If a sentence or paragraph is repeated later in the text, the tool flags this so that the translator’s own wording can be reused in all instances, if the translator wishes. This is very helpful when translating texts in which the same wording is repeated multiple times.
What is machine translation and how does it work?
Another useful function for translators in a CAT tool is machine translation. The tool automatically translates the text sentence by sentence and the translator can then manually edit the translation.
Machine translation is so-called rule-based translation. This is a method based on pre-defined language rules, dictionaries and statistical probability. Although certain types of texts can be translated relatively successfully, this method has its limitations.
The machine translation module is not always successful at finding the correct word in a certain context based on the meaning of the word in the source language. Idiomatic phrases and cultural nuances can also result in incorrect machine translations. Actual spelling errors are uncommon but the translation can be illogical, too formal or lack any natural flow. It can even be grammatically incorrect and contain syntax errors or other inaccuracies.
Because a machine translation may initially look fine, both skills and experience are required to spot errors and edit the translation to make it correct and reader friendly. In many cases it’s better not to use the machine translation function because it can result in a disproportionately long and arduous post-editing process.
When it comes to complex texts or where accuracy is particularly important, such as in financial, legal and medical contexts, machine translation is often not much help at all. The same applies to a text full of nuance that is meant to evoke emotions or convince the reader, –such as in fiction, certain types of marketing texts or a CEO’s comments in an annual report. In these cases many professional translators instead opt to rely on their own skills and translate without activating the machine translation function in their CAT tool. They can quite simply work faster and achieve a better result this way.
How does translation with the help of generative AI work?
AI translation is not rule-based in the same way as traditional machine translation. Instead the system leverages machine learning and deep learning techniques to analyse and understand language in a more profound way. AI models are trained in vast amounts of text data, enabling them to identify patterns and contexts.
So-called neural machine translation (NMT) systems are used. These systems break down a sentence into its various components and then reconstruct it in the target language. NMT models are trained in vast amounts of parallel texts (i.e. texts with identical content in multiple languages). By analysing these texts the models learn to predict how to translate words and phrases in different contexts.
Simply put, generative AI translates from a probability perspective. It relies on statistics and provides responses based on probability.
What does self-learning AI mean?
When people talk about AI being self-learning they’re referring to a process called self-supervised learning. This means that the AI model is programmed to develop its skills without human intervention. Analysing large amounts of data allows the AI model to discover patterns, make assumptions and improve its performance over time.
Can AI translation be trusted?
There is no doubt that AI translation is both fast and, for the most part, good. The latest versions often translate better than traditional machine translation. By how good is it? Despite the many benefits of AI translation, there are also pitfalls. Here are some things to keep in mind:
- AI repeats errors. If its training data contains incorrect information, this can be repeated in translations. This risk is lower in traditional machine translation because there are fixed language rules and dictionaries that have been entered into the model.
- AI can miss cultural nuances, which can result in misleading translations. This also happens with traditional machine translation.
- AI “hallucinates” (yes, this is a common term used in AI research). If AI does not have sufficient relevant information from its training, it may simply guess the answer or fill gaps with information that is not correct. Because the information is presented in a credible way and appears to make sense, it can be difficult to detect errors or inaccuracies.
Don’t forget confidentiality!
It’s also very important to consider confidentiality. Certain information is sensitive and confidential, and it’s therefore important to make sure that the tool being used maintains high standards of data protection and privacy.
Novoterm’s approach to this developing technology
At Novoterm we are monitoring technological developments, constantly educating ourselves, and regularly testing and evaluating new tools and AI solutions. We then apply this knowledge in our translation work to select the methods that are best suited to each assignment, text and customer. Our goal is to always deliver excellent translations to make sure that our customers are satisfied.
Novoterm offers translation into more than 40 languages. We always match the customer’s needs with the right translator and expertise – and the right technical solution.
Get in touch and we will help you!
Novoterm is a highly regarded translation agency that was established in 1996. We offer language services in more than 40 languages provided by translators and proofreaders who understand your industry – we listen, we understand, we translate.