By Viktoryia Baum
Over the past year and a half, the world has been thrown into the hype (or pain?) of artificial intelligence, with the advancement of many products, and the race between the companies to create their own, unique and individual intelligent chatbots capable of many ordinary human tasks. I have been hearing from many colleagues of different ages how this is becoming an existential threat to the profession as a whole, for both translators and interpreters alike. But the real question remains: is it?
I’m a firm believer that if there’s a will, there’s a way. Meaning that I don’t think any of these “robots” will eliminate the need for the profession and the professionals, or somehow bring about our total and ultimate demise. If anything, they can be trained to help us in any imaginative way or method possible. Such opinions of mine have been met with frustration and disbelief by many of my colleagues, yet welcomed and shared by another many.
Recently, Eugenia Tietz-Sokolskaya invited and encouraged me to consider writing for the SLD blog on any subject and anything at all (thank you, Eugenia!), and although I don’t do much writing in any capacity, I figured I might as well give it a try. I usually have a lot to say about things, so why not do it here? Hopefully, the readers will not have a plate of rotten tomatoes handy to smash against the monitor as they read these lines. Maybe, just by sharing some of my experiences I may help alleviate some technology fears, or help someone learn something new. I will probably anger some colleagues and make them reach for their plate of rotten tomatoes, but that’s where technology is in my favor—unlike 200 years ago, I don’t have to deliver this speech in front of you (thank you, industrial civilization!), and I can hide behind the screen.
Let’s start with some basics and try to break down the concept of generative AI, its purpose, and what it can do. Basically, generative AI refers to artificial intelligence systems capable of generating new data and producing content. It can create not just text, but also images or even audio and video files. Here are some examples of generative AI:
- Language models (GPT-4) that can generate human-quality text on almost any subject based on a prompt
- Image generators used to create novel images from text descriptions (DALL-E 2 and Stable Diffusion)
- Music generators that can compose new songs, musical compositions, or other audio using training data (MuseNet)
- Video generators that can make video clips by processing and synthesizing existing data (Sora)
The capability behind generative AI is machine learning models, especially large neural networks. They can analyze thousands of datasets and learn the patterns and representations within that data. Having learned and captured those patterns, the models are then capable of generating new content that is similar to the training data statistically, yet differs from it in a new way. One type of machine learning model is large language models (a phrase everyone in the industry has been hearing more and more); those are built on neural architecture, and they are referred to as large because they use huge datasets. There has been a boost and a rapid increase in research and development of large language models in big tech, with any major player buckling up and racing to develop their own models and tools. The now very infamous ChatGPT is a large language model, or LLM. In simple terms, this is a computer program that has received and analyzed enough data that it is now capable of generating its own responses. The quality of the responses always depends on the quality of the datasets the LLM has been fed. The more they know, the better they are at producing content.
I’ve been experimenting with AI chatbots for quite some time now. When I said they can be trained to help us in any imaginative way or method possible, that’s because I’ve tried and largely succeeded on some level, achieving the objective I set out for myself. A few months ago, for example, I chatted with one of the publicly available chatbots (not ChatGPT) about building up a glossary specifically for use in interpretation, and I’d like to share the outcome.
The background for this experiment was fairly simple. I wear both hats of being a translator and an interpreter. In my state, I am a certified per diem court interpreter, but I live in an area with not a lot of need for court interpreting in my language combination. If anything, it’s approximately 4-5 calls a year plus a few depositions. This means that enough time passes between court appearances that my skills and language databanks typically grow rusty and need a solid refreshment before assignments. Preparing for any court appearances can be daunting due to lack of information and/or materials. If all I know is that tomorrow I am heading over to family court, how should I prepare? What should I look for?
I started with a relatively easy prompt, asking my new AI friend (let’s call him Sam) to help me put together a Russian to English legal glossary to use for interpreting purposes. In a few seconds, the system produced a glossary with some common legal terms, roughly 20 to 25, including terms like lawsuit/claim, indictment, defendant, and court order. The chatbot then asked me if I wanted to know any other specifics to be added “for my translation purposes,” to which I happily said yes, also requesting a set of legal terms in Latin, giving an example prompt of “amicus curiae.” The result was a compilation of another 25-30 terms including bona fide, nolo contendere, ex parte, pro se, inter alia, etc.
For the third part of the conversation, I congratulated Sam on providing me with the definitions of Latin legal terms in the English language, then asking if he could take these same Latin terms and give me their Russian equivalents. He obliged. It is important to note here that the glossary was provided to me in the LATIN to RUSSIAN variant, although for some reason I expected Sam to use the English translations he pulled for the Latin and only then give me the Russian equivalents. I was wrong. Sam was smart, even if my prompt was quite poorly written.
Next, I asked for more terminology, the more advanced, the better. The resulting table included terms like defamation, pre-trial agreement, acquittal, and statute of limitations. I felt that it could have been a bit more advanced, but ok. The next query asked for terms used in family and traffic courts. Then I made another query, and another, and another. But I think I made my point already. Perhaps the best glossary (in my humble opinion) was produced when I asked for terminology specifically used in an arraignment hearing. This turned up terms like request for leniency, motion to suppress evidence, own recognizance release, and recusal motion.
The entire encounter and the prompts along with their results took me approximately 5-7 minutes. The end of the conversation was quite comical, since I asked for all of the terms to be exported to Excel and sent to me via email. Sam happily agreed, took down my email address, and nothing happened. When told that no email had come, he apologized profusely for letting me down and leading me to believe he had the capability of exporting and sending emails. He said that by design he actually did not have that capability, but in his inherent desire to please me, he misrepresented his abilities. The apologetics went on for a few more rounds, much to my amusement and comic relief. It was clear that I wasn’t speaking with a human, but I got amused as if I really were.
Overall, I found the quality of the glossaries produced by AI very high. I did not find any incorrect translations, nor did I find any inconsistencies (some terms were repeated with each of the prompts, and the output was the same). If you ask me personally, I believe that for a real assignment, this approach would have saved me hours of googling terminology and trying to think about all the possible terms I could encounter. Of course, if I were unfamiliar with any of the terms or translations produced, I would have double-checked everything and cross-referenced items using regular old-fashioned dictionaries. For example, if I were to go and interpret at a mining conference, or any other completely foreign subject, like rare and valuable gemstones, I would follow up on what the AI told me.
I forgot to mention that I went into this experiment with very low expectations. I can’t say that I wanted the AI to give me erroneous translations of the terms, to prove that humans and only humans are capable of creating a glossary, but I did think there would be errors. AI proved me wrong, at least for that exercise. I intend to keep practicing and conversing with it regarding my needs in the profession. If you are given the tool, why not use it? And if you are skeptical, I would invite you to try for yourself. Or you can reach for that plate of rotten tomatoes……just kidding. In short, you don’t have to love it or hate it, endorse it or promote it. It’s just a tool. If it can help you save time and do a better job, why not. If anything, you will definitely get some comic relief from experimenting.
If you have any comments or feedback regarding this blog post, I invite you to reach out directly to me via email at vbaum00@gmail.com. In the meantime, I’ll work on Part II of this series over the next weeks. Thanks for reading, and stay tuned!
Viktoryia Baum is an avid technology buff and skilled researcher who spends at least part of her spare time studying new and existing language tools. Having started as an aerospace interpreter many years ago, she discovered she was equally good at translation, so she does both, working with Eastern European languages and specializing in technical, legal, and medical translation and interpreting. She lives in upstate New York and can be contacted at vbaum00@gmail.com.