Quality Insurance: Leveraging AI for the Best Machine Translation

Last Updated May 12, 2023

Accelerate projects and reduce costs by leveraging AI for the best machine translation process.

Generative Artificial Intelligence is already a prevalent device in the translation industry. It helps language service providers finish projects faster and creates more cost-effective processes and workflows.

Still, even the best machine translation output requires human post edits.

Why? Because AI doesn’t quite yet offer an optimal final translation. It’s a great evaluator, though, and it can get you to the final quality assurance phase faster.

Let’s take a deeper look at where we’re at with leveraging AI for the best machine translation process.

3 Traditional Types of Machine Translation

For a while now, there have three types of computer-aided translation systems:

Rule-based machine translation (RBMT)

RBMT requires a full vocabulary and language rules to function properly.

That information dictates how a word or phrase in the source language will translate in the target language.

Basically, the rules state “this means exactly that” in the target language and there’s no deviating from it.

Since language is dynamic and evolves over time, RBMT limits efficacy. A lack of on-the-fly adaptability lends itself to more technical translation projects where set terms remain in place for long periods of time.

Statistical machine translation (SMT)

Here, extensive analysis of both the source and target languages generates statistical models that translate text from one language to another.

You translate the source material based on the most common previous translations.

There’s some nuance, but probabilities rule the day. So, like RBMT, SMT’s general weakness is that it can only translate a phrase if it exists in the reference texts.

Neural Machine Translation (NMT)

Here’s where AI begins to shine.

NMT can recognize patterns in the source material to determine a context-based interpretation that can predict the likelihood of a sequence of words.

While it sounds a lot like SMT, here the machine learns from each translation task and improves upon each subsequent translation.

And here’s the kicker: it doesn’t require human supervision. The linear logic of traditional computers makes way for a neural net method modelled on the human brain. The software learns and perfects after each new experience.

In the end, translations based on neural machine networks perceive the task and consider the context. As a result, these translations are often much more natural than those based on rules or probabilities.

So, machine translation has reached a near-human level of accuracy.

It’s with a Large Language Model (LLM) that we can take things up another notch.

The Large Language Model (LLM)

A Large Language Model (LLM) is a type of AI algorithm that uses a combination of deep learning techniques and massive data sets to understand, summarize, generate, and predict new content.

Think ChatGPT as today’s prime example.

LLMs are a type of generative AI specifically built to help generate text-based content.  The LLM can understand and recognize relationships and connections between words and concepts using a self-attention mechanism.

Once the training has been undertaken, a base exists on which the AI can be used for practical purposes. By querying the LLM with a prompt, the AI model inference can generate a response.

We’re no longer dealing in rules and probabilities, then. And the use cases are great in number.

  • The ability to generate text on any topic
  • Summarizing blocks or multiple pages of text
  • Rewriting content
  • Classify and categorize content
  • Sentiment analysis
  • Conversational AI and chatbots
  • Translation

We’re written extensively about ChatGPT’s capabilities and limitations, and specifically with regards to translation.

We’re not yet at the point where translation can be automated with 100% accuracy, meaning a human touch is still quite necessary. But AI can certainly expedite the process when appropriate.

Here’s how.

Using AI to Evaluate Translation

AI models simply haven’t risen to full, accurate translation capabilities.

You can certainly use it for extremely fast, high-volume translations at very low cost. But it comes at the risk of low accuracy and the possibility of critical errors.

For technical, marketing, financial, legal, or medical content, the slightest error or ambiguity can have damaging consequences, generate disputes, and have a major impact on your company’s reputation.

And even with respect to website or e-commerce content, there are certain risks, especially in target markets that are further afield.

Microsoft compared ChatGPT to more traditional MT engines in 18 high- and low-resource language pairs. They used publicly available datasets and tested at both sentence- and document-level.

The conclusion was GPT models produce a “very competitive translation quality” for high resource languages. However, ChatGPT still has “limited capabilities for low resource languages.”

For reference, English, Chinese, Spanish, French, Japanese and more of the European and Western languages are high resource.

To be more successful with low resource languages, text data collection and annotation is necessary. If you’re translating into these languages, ChatGPT won’t be enough to get the job done, and that’s for basic content.

Having said that all, AI will and already does undoubtedly facilitate the work of the translator.

NMT + LLM + PE

We are already seeing the emergence of a hybrid method where AI is combined with human know-how: neural machine translation with large-language model evaluation and human post-editing.

Here, AI helps mostly at the project management level:

  • Explain the quality of a translation.
  • Pick the best translation out of two options.
  • Identify and explain serious mistakes in translation.

We can’t trust it to spot all mistakes, and some mistakes maybe aren’t important depending on the scope of your project. The mistake spotting is already clever, though, and the way it criticizes is dependent on your instructions.

You insert the prompt and the context, and the AI will work to verify whether your NMT is up to par.

Again, we’ve found AI is best for EVALUATION as a quick step between MT and PE that can save both time and money without costing you in terms of quality.

Run it through the machine translation process, test accuracy with AI, then comb it over with post-editing by humans.

The Best Machine Translation Built For You

There’s no one way to approach translation projects. We’re big on flexibility around here and offer the specialized services to back it up.

It’s a sliding scale between pure human translation and handing everything over to the machine.

Human translation is slow but solid, and it comes at a cost. The quality, though, is premium.

Machine translation tools make things quick and simple, and you can get it done inexpensively. The quality, as we’ve seen, isn’t outstanding.

But with AI evaluation and post-edited machine translation, you can achieve a greater semblance of balance.

We consider all the factors before rolling out a custom price.

First and foremost, you must know what exactly you need. Make a list of the languages into which or from which you need to translate and decide on the level of language services your company needs.

The more information you provide to the person quoting for your translation needs, the greater the chance that you will be offered the most attractive price.

And if you don’t know exactly what you need, we can help figure it out.

Get in touch to get started.

Related Posts

Summa Linguae uses cookies to allow us to better understand how the site is used. By continuing to use this site, you consent to this policy.

Learn More