We often present clients with guidance on how to work with interpreters, and frequently get asked about AI. This is because many people are waiting for the day that they can simply go online and use AI to seamlessly translate between two different languages, but we would like to say it out loud here: THAT DAY HAS YET TO COME.

Please also check out this flip-book we’ve made https://heyzine.com/flip-book/20de67a12a.html

Please also check out our playlist for Chinese localization case studies: https://www.youtube.com/playlist?list=PLO-QGEbwcTr14xqfiR38Mp-EhHAmclsUY

We localized the Interpreters and Music video into traditional Chinese as an example to compare translation accuracy between humans versus AI and to identify some classic AI issues.

One of the biggest weaknesses of AI is that it often struggles with names. For instance, the name “Laura” was translated into both “蘿拉” and “勞拉.” When we saw this inconsistency in names, we looked at each other with amusement because this happens all the time. Some may say AI spelling names incorrectly isn’t a big deal since it’s an easy fix. However, for those people, we’d like to share a real-life example.

In a lease contract we worked on, Paragraph 1 said that the landlord shall be known as ‘A’ and the tenant as ‘B’. Paragraph 2 called the landlord ‘C’ and the tenant ‘D’. This was a document with 30,000 words that a client asked us to quote for reviewing the translation, which had probably been done by an AI. Just in terms of reviewing names, how much effort would it take to find out if there were places that call the landlord “E” and the tenant “F” and so on? Not to mention all the work it would take to find other mistakes that humans typically need several rounds of review to detect.

Our analysis also uncovered that AI defaults to using the pronoun “你,” referring to males and offering no female form “妳.”

AI have translated love song titles like “Suddenly Missing You” and “Stuck on You” into traditional Chinese, using the male form. The male singers may not prefer using the male form of “you” in their love song titles. Otherwise, a native speaker in traditional Chinese would feel kind of strange, reading it.

We inserted line breaks on messages that appear in the video. With line breaks, AI seemed to lose the context of the lines.

Line breaks are important. We are often requested to insert line breaks in Asian language marketing materials. Take Japanese line breaks as an example. There are some basic rules for line breaks or how to break words up, but at the same time, there are a lot of exceptions, which humans can easily catch if they understand Japanese, but not AI. In other words, humans break things apart (debriefing) and put them together in a creative way, which AI is just not capable of.

It turns out that AI struggles to translate any segment accurately and, at times, produces unnatural and contextually absurd translations. As shown in the screenshot below, even with a relatively short source text, the quality of AI translation was unbelievably subpar.

AI translated “interpretation” as “explanation” due to a lack of context.
AI translated “Performance” to machine’s performance rather than that of the interpreter’s.
AI mistakenly translated the meaning of “like” as “to be fond of” instead of “similar to.”
AI word-for-word translation for “big heart” doesn’t make sense to a Chinese audience.

It’s clear to us that AI is not able to handle messages that are broken down by line breaks. This then leads us to a question: How well could AI handle entire messages without line breaks?

We conducted a retest by removing all the line breaks on messages. In this attempt, the text was formatted in a more machine-friendly way to enhance AI’s understanding. But even so, post-editing remained an essential step, with 80% of the segments requiring significant human intervention. Without this crucial step, AI translations either come across as rigid and less relatable to our audience, or contain mistranslations. Below are some examples.

Example 1:
The AI translation appears rather stiff because the word “sync” was translated literally. The audience might wonder what it means to “sync” one language to another. Human translators are able to further explain the context of sync, that is, interpreters “listen to one language and convey it in another language.”

Example 2:
AI translated “more emotionally acute” as “more impatient,” which not only deviates from the intended meaning of the source, but also negates the impact of the word “music”. During post-editing, we replaced it with “more emotionally sensitive,” which is more contextually accurate.

Example 3:
AI did word-for-word translation again. It doesn’t sound like what a normal person would say in Chinese. As a dynamic language, Chinese favors verbs over nouns and usually keeps sentences short. Therefore, in post-editing, we restructured the sentence to make it fit a typical Chinese writing style, and flow more naturally.

Example 4:
AI’s translation of “concentration” lacked clarity. Without referring to the source, it was hard to grasp the intended meaning. So, we opted for a more precise choice of words.

Example 5:
AI does a literal translation, full of ambiguity and rigidity, which doesn’t make clear sense to a Chinese audience.

Translation is supposed to flow naturally to engage the audience. It is the more immersive and relatable experiences that make humans feel comfortable. These are exactly the areas where we as interpreters and translators can contribute to.

There may be a lot of gloom and doom from some in the community who think that their jobs are at risk, however, the reality is that we’re training AI to speak our language, but they aren’t able to fully understand it like we can. They can process it, try and find the corresponding pattern in their database, and come to a conclusion that they think is right, but they won’t always be. That’s where interpreters and translators will always have the edge over AI. Human creativity and our ability to understand what’s important, and the culture embedded in it, enables us to make sure that we are conveying the intended message.

We tried one of the latest AI platforms to translate one of our office videos into Mandarin.

While we were impressed by the seamless process and the voice cloning feature that enhanced voice modulation, we couldn’t help but notice pronunciation and translation errors in the generated video. Given that Mandarin Chinese is a tonal language, tones can become a source of misunderstanding if not pronounced correctly. The chosen video introduces the rental service of our conference room, making “conference” a high-frequency word. However, throughout the video, AI consistently pronounces the Chinese word for “conference,” as “memory,” with tones differing from the former. Also, “state-of-the-art” in Chinese is pronounced the same way that “cash” is. This could undoubtedly complicate the message we aim to convey if left alone.

The translation issues we caught are mostly recurring problems caused by machine translation as discussed above. Take the first sentence as an example. AI translated “Looking for a conference room to have a meeting over video or in person?” as “Can you look for a conference room via video or in person meeting?” AI’s rendition deviates from the original meaning, which is likely caused by line breaks, leading to confusion and miscommunication. Such discrepancies underscore the importance of post-editing and human intervention to refine machine-generated translations.

Our conclusion becomes clear.

In this age of AI becoming more prevalent, humans just need to work smarter to beat out AI. As individuals in an evolving world, it’s important to accept technological advancements, but also understand that AI lacks creativity, individuality, improvisation capability, and the understanding of human cultures. That’s how humans can break through and go beyond AI’s limitations.

Video link: https://youtu.be/dATBteNQ-zY?si=bgPDXw-3FblyfnaW

Picking up where we left off on our previous behind the scenes diversity blog, we are certain that there is one thing that AI has yet to really accomplish. AI’s inability to generate something unique in music and in interpretation is something interesting to explore.

We have started to show the video, Interpreter and Music, to our closer circle of colleagues and friends in order to gather feedback. One colleague sent us a video from Kuban Music on YouTube about AI singing a Chinese song. Our initial reaction was of surprise at how impressive it is that an AI was able to sing like that. Our director of translation Mei-Ling thinks it’s so cool that she started to learn the song from the AI. But the more we listen to it, the more we begin to realize that there’s something lacking in its composition and performance. After repeated listening, Mei-Ling, who is not even close to be called a singer, feels the urge to record herself to challenge the AI singer. She said it was a therapeutic experience, because she feels that she has something unique that the AI is lacking — heart and liveliness. There is something special from hearing a human singing that an AI voice couldn’t produce. For comparison, please see both the AI and the human samples below. We are sure if we were to invite 100 humans to sing, they will come out all differently with their own individual uniqueness, while AI variations will most likely be limited.

Everyone knows that AI is a result of big data, “aggregating” from others to generate its materials. It’s trained in a way to mimic humans. Notice that’s humans, plural. This is because it’s not trained to mimic individual people, but a general populace. It essentially crowdsources behavior from the resources it’s trained on, and becomes not just one person, but many. Though the programmers may have the AI come up with clever phrases, such as AI shedding tears, to make it sound like its own person, it isn’t nearly as unique and diverse as humans. It’s inherently impossible for that to happen because of the origin of the AI’s nature.

On the contrary to that, with our interpreter and music video, individuality is one thing that we really wanted to express. 15 individual contributions from around the world bring in incredibly vast diversity and inclusion, yet no interpreters were instructed on what exactly to be contributed. Every piece comes out of interpreter’s individuality with each interpreter’s own language, culture, thoughts, feelings, circumstances and education that leads to how they act. In this sense, everyone is as unique as a fingerprint—there really is nobody like each one of them in the world.

For music, it’s always about expressing oneself with human emotions and energy, and AI may be struggling to do that because it doesn’t have a sense of self. The ability to have your emotions flow out is something that AI is not able to replicate, at least not in present times. AI is always searching for the “right” answer, but there isn’t one with creative media like music or abstract concepts like human energy.

AI is unable to think outside of the box and improvise like humans can. They’re trained to think on pre-existing thoughts and opinions. Anything outside of that scope is not something that they are able to even perceive. That’s why, for example, AI will never truly be able to replicate jazz. Jazz is a collaborative performance where innovation and creativity are key elements. At any given minute, the tempo and melody can change to something entirely different, and it’s up to the musicians to keep up and complement each other so that it doesn’t sound like complete chaos. AI is very exciting technology, but there’s nothing quite as thrilling as watching jazz musicians collaborate or excellent interpreters work.

More often than not, interpretation is similar to Jazz where interpreters often need to improvise and come up with creative solutions to their on-the-spot translation. AI is able to follow patterns to come up with solutions that others have thought of and have done before, but they’re not able to come up with these types of originality their own. AI singing or interpreting at the current stage may serve as a gist of ideas, a prototype, or a machine for learning a song. Still, it will be important for humans to intervene, and add the kind of individuality or human touch that will help it sparkle and make it shine and standout as a piece of unique individual work.

In the future, the methodology used to create AI may change. Is it possible that AI will get closer to having more individuality and uniqueness? We’d love to hear our colleague’s thoughts on the topic because everyone brings their own unique perspective. There are multiple answers to this question and we’re happy to start the discussion and see where it leads us since this is such an important topic in this day and age!

AI singing sample: (Listen Here)

Human singing sample: (Listen Here)