Generative Translation: Correcting undesirable behavior
27 Aug 2024
8 mins read
Before I get too deep into this post, I’d like to make it clear that my children are (normally) well-behaved. And so they should be; they’re 16, 20, and 21 now. However, this hasn’t always been the case and many times over the years we’ve had to “gently correct” (make of that what you will!) how they speak, which words they choose to use, and set a good example of how to generally be a better person; constantly guiding them in the right direction so they know better next time around. What’s this got to do with translation? Good question! Let me take you on a trip down memory lane.
Recap
Back in September 2023, I wrote an article about how AI was going to turbocharge Translation Engines in Trados. Back then, I referred to Chicken Run and likened our Translation Engines to a pie-making machine that was the nemesis of many an unsuspecting chicken. Instead of chickens, we were feeding source text to our engines. Instead of pies, we were getting translations.
Not long after my original post, we launched our Generative Translation capability which uses AI to either improve an existing translation or translate text from scratch. Our goals were:
- Improved quality of translations for post-editing.
- Increased productivity through reduced post-editing effort.
- Faster time-to-market.
All of the above will hopefully lead to a better end product and higher customer satisfaction. Everyone in our industry is curious to see whether AI, when given the right prompt, can enhance or even replace existing Neural Machine Translation capabilities. Right now, I don’t think we’ve got a definitive answer to the question. We have, however, been able to evaluate the results and improve on the quality of the translations using various approaches:
- Adjusting the prompt that we give to the LLM to produce the translation. (Side note: Who knew five years ago that “Prompt Engineer” would be a job title?)
- Providing the LLM with terminology to use in the translation.
- Using previously translated content to determine the style and tone of the translation.
- Allowing users to provide custom instructions to the LLM.
Correcting “bad” translations
So then, let’s revisit those children. Suppose I want to translate the following phrase from Spanish into English:
Mis hijos son muy traviesos.
Let’s try asking an LLM to translate this. Here’s a sample prompt that I engineered (can I have a job?):
I want you to translate some text from Spanish into English. The text to be translated is surrounded by <<< and >>>. Your response should only be the translated version of the text in English.
Here is the text:
<<<Mis hijos son muy traviesos.>>>
And here’s what we get:
My children are very mischievous.
I’m a funny soul. I prefer to refer to those little angels as “my offspring”. Also, “mischievous” is not a word I would use outside the context of Halloween, so I’d rather use “badly behaved”. This is my own terminology, and I’d like to have the LLM use that instead. How about this prompt:
I want you to translate some text from Spanish into English. The text to be translated is surrounded by <<< and >>>. Your response should only be the translated version of the text in English.
I want you to translate "hijos" as "offspring".
I want you to translate "traviesos" as "badly behaved".
Here is the text:
<<<Mis hijos son muy traviesos.>>>
Now I get:
My offspring are very badly behaved.
Perfect! Well, it would be if they were better behaved, I’m sure. Suppose I want a distinctive style of translation because I’ve had something translated before that I like. Here’s another prompt:
I want you to translate some text from Spanish into English. The text to be translated is surrounded by <<< and >>>. Your response should only be the translated version of the text in English.
I have a previous translation style that I'd like to copy when "Mis hijos están muy felices." in Spanish was translated as "I have very happy children." in English. Make the translation match this pattern.
Here is the text:
<<<Mis hijos son muy traviesos.>>>
Now I get:
I have very mischievous children.
I guess this isn’t really something to be proud of, but suppose I want to shout about those cheeky little imps from the rooftops so I’m going to ask something else of the LLM:
I want you to translate some text from Spanish into English. The text to be translated is surrounded by <<< and >>>. Your response should only be the translated version of the text in English.
Make the translation in all caps.
Here is the text:
<<<Mis hijos son muy traviesos.>>>
And the response:
MY CHILDREN ARE VERY MISCHIEVOUS.
We’ve seen how we can use terminology, previous translations, and custom instructions to get different results from the LLM. The culmination of all of this would be to combine everything into a single prompt:
I want you to translate some text from Spanish into English. The text to be translated is surrounded by <<< and >>>. Your response should only be the translated version of the text in English.
I want you to translate "hijos" as "offspring".
I want you to translate "traviesos" as "badly behaved".
I have a previous translation style that I'd like to copy when "Mis hijos están muy felices." in Spanish was translated as "I have very happy children." in English. Make the translation match this pattern.
Make the translation in all caps.
Here is the text:
<<<Mis hijos son muy traviesos.>>>
And we have:
I HAVE VERY BADLY BEHAVED OFFSPRING.
Not something that I really want to shout about but, then again, as previously discussed, this isn’t actually true. Perhaps I should try again and translate “hijos” as “angels”?
What does this mean for me?
Our Generative Translation feature uses all the techniques described above. When included in your project workflow, Generative Translation will use the project settings, including the Translation Engine and Custom Fields, to do the following:
- Terminology: Find terms from terminology databases in the source text and send them to the LLM.
- Translation Memory: Find matches from the translation memories and provide them to the LLM as example translations.
- Custom Instructions: Retrieve specific instructions defined in the translation project, like target audience, tone, or length restrictions, and include them in the prompt for the LLM.
Using these techniques means that your LLM is guided to provide better translations, reducing your post-editing efforts which means cheaper translations and faster time-to-market.
Accessing Generative Translation
Customers who want to get their hands on Generative Translation can do so by purchasing the AI Essentials add-on for Trados Team, Trados Accelerate, and Trados Enterprise. Contact us today to find out more.
Alternatively, if you’re a translator using Trados Studio, you can also get generative translation capabilities by using the new AI Assistant in Trados Studio 2024.