I recently attended a trade show where the owner of a training company proudly boasted how all of their training content is available in 100 languages. She said that Google Translate powers the translation in her training content. She discovered the single step that makes the content 95% accurate: use the active voice, and you are all set. Sadly, she is 100% wrong. In all honesty, this is disingenuous, dangerous, and flat-out deceit.
Any translation effort you engage in is about managing the cost and the risk of errors in your content. Think about it. Simply flipping the switch on a service like Google Translate or any other AI-enhanced workflow and using it with high-risk content, like the software and instructions for a defibrillator, would be criminal.
Machine Translation is powerful. It is a tool that can be effectively used for customers looking to break down language barriers efficiently and inexpensively. However, simply using the active voice will not give you the claimed 95% accuracy.
Properly using AI, neural machine translation (NMT), and human verification in your workflow can accomplish a suitable result by minimizing your risks and costs. This is possible and encouraged by many of the more forward-thinking language service providers.
For those interested in understanding how to incorporate AI translation effectively into their workflows, our AI translation eBook offers practical guidance on making informed decisions, including service levels, customization, and human involvement.
Training a neural machine translation engine
Implementing AI translation successfully goes beyond just using off-the-shelf tools—it requires properly training a neural machine translation engine to meet your specific needs. Here’s a breakdown of what that process involves:
- Translation Collection and Preparation: The first step involves collecting a large corpus of text that contains human-verified, parallel sentences in both the source and target languages. This text is then cleaned and preprocessed, which may include tasks like normalizing punctuation, removing noise, and splitting the text into manageable segments. Most services recommend a minimum of 10,000 segments for this process.
- Model Architecture Selection: An NMT model typically uses a sequence-to-sequence (seq2seq) architecture, which is composed of an encoder and a decoder. The encoder reads and encodes the source sentence into a fixed-length vector, and the decoder then generates the translation from this encoded vector, one word at a time. Recently, transformer models, which use self-attention mechanisms, have become the standard due to their effectiveness and efficiency. Some of the factors going into which model you select will be the languages you plan to use and the technology your team prefers. There are many choices on the market, such as Azure, DeepL, Hugging Face, etc.
- Training Process: During training, the model is shown the source sentence and is trained to predict the corresponding target sentence. This is done using techniques such as teacher forcing, where the correct output (target sentence) is shown to the model during training to improve learning. The model’s predictions are compared to the actual target sentences, and the difference (error) is calculated using a loss function. The model’s parameters are adjusted to minimize this loss through backpropagation and optimization algorithms like Stochastic Gradient Descent (SGD) or Adam. Most of the popular NMT engines allow for training and are relatively inexpensive.
- Evaluation and Iteration: The model’s performance is periodically evaluated on a separate validation set that was not seen during training to monitor its translation quality and to prevent overfitting. Metrics such as the BLEU (Bilingual Evaluation Understudy) score are commonly used to quantify translation quality. The training process is iteratively refined based on these evaluations. The process can also include further human verification and subsequent retraining of the NMT engine.
- Fine-tuning: In some cases, the NMT model is further fine-tuned on a smaller, domain-specific dataset to improve its performance on particular text types, such as technical manuals or other forms of user documentation.
Final thoughts
The bottom line is that creating a proper NMT engine will yield significant improvements over simply using out-of-the-box machine translation. The hype will tell you that simply adding Machine Translation to your workflows solves all your problems. The real truth is that including a human in your translation training and subsequent verification is the true path to optimizing your AI-enabled process, resulting in a better outcome.
