Understanding Neural Machine Translation
Since its inception in 1954, machine translation has come a long way.
The technological revolution has seen machine translation (MT) evolve, from largely inaccessible, labour intensive, time-consuming computing processes, to more sophisticated neural machine translation (NMT) models, which are widely used to support the day-to-day business communication of organisations across the world.
At Andiamo, NMT is our chosen machine translation model, and in this blog we tackle how and why it’s used, its benefits over those of other MT models, and more.
How Does Neural Machine Translation Work?
NMT is used to help LSPs to accurately translate content from one language to another, in a cost-effective way that supports quicker turnaround times.
These models use artificial neural networks (a form of AI machine learning based on the workings of the human brain) to recognise patterns, predict word sequences, and translate content in sentences, rather than word-by-word as old machine translation models did.
They work by passing information through their networks of interconnected nodes to create a translation.
Neural machine translation models are trained on datasets of translations (which can include your own previous translations), and learn by recognising the patterns in them, as humans do. They’re quick to learn, and learn continuously, but the quality of the output is very much dependent on the quality of the bilingual text used for training them.
NMT vs Other Machine Translation Models?
There are three kinds of MT: rule-based, statistical, and neural machine translation. Here’s how they each work, broadly speaking, and why we believe NMT is the ultimate option.
Rule-based MT: RBMT is a time consuming model to use, as the rules of each language have to be manually input into the system. This less sophisticated, manual mode of operation is an inefficient way of training a machine translation model, as the rules of language evolve over time. Furthermore, RBMT models often produce awkward sounding output.
Statistical MT: SMT is a more limiting and perhaps outdated MT model that works by pulling phrases from parallel corpora, which are publicly available datasets of translated texts used to train MT models. If a phrase doesn’t exist in the reference translations, the engine won’t be able to produce a translation.
Neural MT: Neural machine translation has a higher translation accuracy than preceding models due to its self-learning capabilities. This means the output needs less post-editing than other MT models, helping to lower service costs.
The Benefits of Neural Machine Translation
- NMT models generate quality, reliable translations of non-minority languages (minority languages are those spoken by a small percentage of a particular territory, such as Gaelic). This means LSPs need to only post-edit the machine translated content, rather than translating the content from scratch, saving you time and money.
- Though we at Andiamo! predominantly use neural machine translation for technical and generic content, it can be employed in a wide range of applications, offering time saving and cost-effective benefits across a range of projects.
- Neural machine translation can be used alongside your existing translation memory (TM), and adopted into the TM software. At Andiamo, this helps to support our human post-editing process to ensure our final translations are consistent with your previous translated content.
- NMT can be highly advantageous for clients who have time-critical projects as it can take a fraction of the time to translate than if a human were translating them from scratch.
NMT Drawbacks and Considerations
- When it comes to training a neural machine translation model, it’s important to only feed the most accurate reference translations into the model. If, for example, a source sentence used for training isn’t correctly aligned with the right target sentence, the output is likely to be detached from the source content which can lead to hallucinations and misleading the reader.
- NMT models are trained on parallel datasets, for example, English > French. Models don’t have the same amount of parallel data in all language combinations, which means their output will be more accurate in the combinations they are more familiar with.
How Andiamo! Uses NMT
At Andiamo, we have a hybrid approach to neural machine translation, because we believe the human touch is always essential.
We use NMT alongside our client’s existing translation memory, so the output is consistent with the terminology saved in the TM. The use of both TM and NMT supports accuracy, saves time, enables us to offer additional discounts to those offered by TM alone, and enables us to use our QA tools to offer greater accuracy, consistency and quality.
Our experienced linguists translate the terms the translation memory and NMT model can’t find the right match for, and post-edit the final translated piece to guarantee its quality and accuracy. At Andiamo, NMT output is always revised or post-edited by a human translator who is a native speaker.
We are always 100% transparent about when we use NMT, and provide quotes that offer pricing options for human translation or NMT (+ human post-editing). With Andiamo, you can be sure that you’re receiving an honest quote for an honest, transparent service. We will never use NMT for a human translation service.
If we can help you with machine translation services, contact our team today to discuss how we can support you with neural machine translation.
