Transcript
Transcript: The GCtranslate Prototype, by Public Services and Procurement Canada
[00:00:00 Text on screen: AI-Powered Projects in the Government of Canada; The GCtranslate Prototype, by Public Services and Procurement Canada.]
[00:00:02 Split screen: Zepporah Toh and title slide. Text on slide: AI at PSPC; Using AI to Improve Outcomes, Practical insights from PSPC's enterprise AI implementation,
Zepporah Toh, Senior Director, Artificial Intelligence and Cloud Service.]
Zepporah Toh: Typically, when we talk about artificial intelligence, the focus is very much on technology, so I'm glad we're talking about the outcomes, so you can actually deal with AI from a business standpoint. Next slide, please.
[00:00:11 Split screen: Zepporah Toh and slide and slide titled: Before You Start: Two Questions to Ask Before Using AI. Text on slide as described.]
Zepporah Toh: So, before you start any kind of AI initiative, and certainly with GCtranslate, you want to always ask yourself — and you've seen it with other presentations today — what is the problem you're trying to solve? And what are the outcomes you're trying to achieve?
And I'll talk about that later in the presentation as well, because very often we're very fixated on the technology. Certainly your CFO will thank you because you get to determine what the return on investment is, but really that should be your core focus. And by asking those questions, if you can articulate that, then you get to turn the focus to value and not so much on technology.
Next slide, please.
[00:00:47 Split screen: Zepporah Toh and slide. Text on slide as described.]
Zepporah Toh: So, about our feature use case today, GCtranslate. Next slide.
[00:00:47 Split screen: Zepporah Toh and slide. The slide, titled "AI is Not New to PSPC's Translation Bureau," shows human evolution, from ape to a caped superhero, as an analogy to the evolution of Automated Translation. Text on slide: ~1977, Introduction of Automated Translation (weather alerts); ~2000, Introduction of translation memories; 2010, Implementation of a centralized and segmented translation memory (Megacorpus); 2016, Launch of a language comprehension tool for federal public servants (statistical machine translation); 2017, Introduction of neural machine translation (NMT) for Bureau translators; 2020, Introduction of trained NMT for Bureau translators, Trained legal NMT 2022; 2025, Analysis and testing of generative AI in the form of large language models (LLM).]
Zepporah Toh: I want to start by saying that AI is not new to PSPC's Translation Bureau. I work with Digital Services Branch within PSPC, and before we actually partnered with our colleagues at the Bureau to make this a reality to roll it across the GC, they'd actually been on this journey for quite some time. They've actually been working with artificial intelligence, and the reason why I'm bringing this to your attention is just because the question is usually asked, okay, if you identified this problem, why did you pick AI as a solution to solve it?
Well, because it's one of the most researched and understood use cases, machine translation, and certainly Translation Bureau has got a lot of expertise, and they've been doing this for quite some time. So [I] just wanted to put that there because this is not new to PSPC's Translation Bureau and certainly something that's already been explored to some degree right now, from neural machine translation to generative AI, which is being used today. Next slide, please.
[00:01:44 Split screen: Zepporah Toh and slide titled: GCtranslate GC Enterprise Translation. Text on slide as described.]
Zepporah Toh: So, what is the problem we're really trying to solve here? With the advent of artificial intelligence, and the internet, and tools available online, employees were getting a little impatient with waiting for translations. Imagine we're sending a request and waiting a couple of days for a text to be translated, or a couple of weeks.
And so, they started using some of these internet tools online, which are not very secure.
Think Google Translate available on the internet. And there's no telling where the data resided, where it was being stored once it was translated, maybe something that is Protected online, exposing the GC to security risk. And on top of that, these tools are not trained with Canadian linguistic data, therefore it's missing that nuance. We all know the Canadian French to English translation is not the same as what you get online.
And so, the solution that we came up with is called GCtranslate. And this is an AI-powered solution that is trained using Translation Bureau's extensive and very powerful corpus of bilingual data, and it's now available up to Protected B, so it's definitely a win other than using internet tools for translation. And we're looking to make it available across the GC2 and north of 350,000 public servants. Next slide, please.
[00:03:08 Split screen: Zepporah Toh and slide titled: Progress to Date, Rapid Growth and Strong Adoption. Text on slide as described.]
Zepporah Toh: And so far, what have we noticed? It was rolled out across PSPC in June of this year, and at the time was a pilot called PSBC Translate. A couple of months later, we rolled it out to 5 early adopter departments and agencies as GCtranslate, and it's been very well received. To date, it's translated more than 100 million words. Actually, we received metrics this morning, it's up to 142 million words to date. And what we noticed is, in just 3 months, this tool was able to translate 4 times the volume that PSPC's Bureau typically handles, the annual volume.
And what is this telling us? It's two things here. First, obviously, you cannot compare manual translations with AI translations. That's the first thing. And secondly, this tool is being used by employees' day-to-day translations, things that never even went to the Bureau to start with. Now they have a tool at their fingertips to carry out this translation. So, it's definitely a game changer. At PSPC, it's one of the highly used applications, so it's made such a difference in our day-to-day work from an official languages standpoint. Next slide, please.
[00:04:17 Split screen: Zepporah Toh and slide. Text on slide as described.]
Zepporah Toh: Okay, so I'm going to talk about 5 key lessons learned, and certainly we have more, but we only have time for 5. And behind this shiny tool is a lot of bruises and tears and sweat that went into it, and things that we had to learn. And I'm sure my colleagues on the screen can share the same, but hopefully they can help you when you're rolling out your own AI projects. So, first lesson — next slide —
[00:04:43 Split screen: Zepporah Toh and slide. Text on slide as described.]
Zepporah Toh: Start small, scale fast. So, if you think about those of us who have been in project management for years now, typically when we're looking to roll out an enterprise project, certainly one of the scope and scale of GCtranslate, we think about putting in this humongous TB submission and mega projects and mammoth initiatives with huge budgets. And, of course, it takes forever to get them off the ground. Before you actually complete the TB submission and see something that folks can use, it takes a lot of time and effort.
And so, fortunately, our senior executives were ahead of the game, and they asked us to deliver an MVP, what you call a Minimal Viable Product. And what that means is the smallest possible functional product that we could roll out that employees could actually use tangibly. And then from there scale it up as time goes on.
Within the case of GCtranslate, we rolled out this MVP across PSPC as PSPC Translate, and then from there we've rolled it out to 5 early adopter departments and agencies. Some of you are already using the tool. And right now, in some kind of a [inaudible] period, operational readiness, we're now trying to fortify the underlying infrastructure, the AI models, to curate more of the data and learn some of the lessons to now inform how we're going to now roll this solution out to the rest of the GC.
So, starting small, we started with a minimal viable product. We didn't build a whole mega solution and then scale first. And the reason why this is very important is because AI is evolving almost at the speed of light, it's changing almost on a daily basis. And so, if you're waiting to build this big mammoth project and roll it out, obviously behind the game. So, this is definitely a key lesson. Next slide, please.
[00:06:27 Split screen: Zepporah Toh and slide. Text on slide as described.]
Zepporah Toh: Next one is good AI requires clean data. And I think it's you, Roxanne, who mentioned that "you eat data for your vegetables because you want to get AI dessert", or something like that. It certainly was the case here.
Behind this tool, there was a lot of time and effort that went into curating and cleansing the data. Fortunately, as I mentioned, Translation Bureau does have a very extensive bilingual corpus of data. But still, before it could actually be used in GCtranslate, there was a lot of time that was required to actually curate that data, to organize it, to clean it up. And we're still going through that process, as well, for the GC-wide rollout.
So, all this to say, and I think everyone has already mentioned that here, that in order for you to actually have a solution that provides very solid outputs, reliable outputs, accurate outputs, especially from a linguistic standpoint, you want to make sure that you're training your AI models with clean data. If not, whatever you train it with is the output you're going to get, and certainly we do not want that. Next slide, please.
[00:07:30 Split screen: Zepporah Toh and slide. Text on slide as described.]
Zepporah Toh: Governance should be an enabler. And this is really near and dear to my heart because, again, my background is in IT project management. And anyone who's gone through a project management lifecycle knows that we have gates and we love our gates. And in government, we have red tape. And, to some extent, I think we're used to it.
But when you're working on an AI project, governance needs to be an enabler. I'm not saying that those checks and balances are not necessary. They're there for a reason; to make sure that we understand the risks, and we're financially responsible, and all that type of stuff.
But sometimes we create all this documentation. We have all these gates to go through and all kinds of steps and committees. And are they, quite frankly, necessary? Do we actually read all of those documents? We can ask those questions. And that's something that we had to go through with GCtranslate. We were asked to roll out this product very quickly, and we did not have the time to actually go through all of those processes, so we had to think and pivot in many ways so that we could actually deliver it in a timely fashion.
So, again, presenting this as a lesson learned because you will run into that. And if you stick with your traditional project management, or product management processes, you're definitely going to end up rolling out a solution in the next 2 to 3 years and we need to move at the speed of AI and not the speed of government. So, that's definitely one lesson we learned. And next slide, please.
[00:08:52 Split screen: Zepporah Toh and slide. Text on slide as described.]
Zepporah Toh: Enterprise AI requires Enterprise IT infrastructure. What does that mean? It's one thing to have your application, the AI application, or your solution, but it has to reside somewhere. Where it needs to function, especially if you're looking to roll it out across the Government of Canada, like GCtranslate.
So, take for example, the solution was almost ready to go, but we did not have the necessary cloud infrastructure yet built. That took some time and approvals to get it in place. And then there's authentication. Obviously, to have a Protected B solution, we needed to have the right secure infrastructure in place. And that, as well, took us some time and effort to put that in place. So, that was a bit of a pain point that we're still dealing with as we're getting ready to roll out this solution across the GC, and we're working with our partners at SSC who are working very hard right now to make sure that we have that in place to enable the GC-wide rollout.
But all this to say, this was a big lesson because it's one thing to have your solution, but if you don't have that Enterprise IT infrastructure, you're not going anywhere. Especially if you're looking to roll out a solution like GCtranslate across the federal government.
[00:09:58 Split screen: Zepporah Toh and slide. Text on slide as described.]
Zepporah Toh: And, last lesson, AI success depends on early Change Management. These tools are going to be used by humans, by employees, and every employee is on a different level when it comes to their AI journey. And so, you have to be mindful of that.
And certainly, we know GCtranslate, as much as it's a relatively simple tool, we had to take into consideration our user base, what kind of communication is required internally to PSPC across the GC, and even folks who are going to use it on a day-to-day basis. There's training sessions as well, workshops that are in place, how to use this tool in a responsible fashion. To understand that you cannot translate secret documentation, for example, on this tool and why it's important to adhere to those restrictions.
And so, that Change Management, making sure you incorporate it from day one is really key because at the end of the day it's a tool that's going to be used by your systems, by various groups within your department, by employees. It's going to change business processes and therefore it's really important that we understand the requirements, educate users, and take them along the journey as we roll out the solution, certainly within PSPC and across the GC. So, that brings me to the end of my presentation. Last slide.
[00:09:58 Split screen: Zepporah Toh and slide. Text on slide:
Looking Ahead: AI is moving fast and already delivering results at PSPC;
Scaling enterprise AI will amplify GC-wide impact; PSPC will continue building shared AI capabilities the GC can reuse.]
Zepporah Toh: Looking forward is just listening to all the presenters today. We all have AI solutions, so it's really important that we collectively continue to share this body of knowledge because we don't want to keep repeating. For example, Gil, you presented AI Answers, and I think I'm going to reach out to you to talk about some of that because we're looking to roll out external chatbots.
So, from a GCtranslate standpoint, we do know that there's a strategy on tools across the GC, so being able to bring all of this intelligence together and work together makes us to be more efficient as a GC.
So, that's pretty much the end of my presentation.
[00:11:49 The CSPS animated logo appears onscreen. Text on screen: canada.ca/school-ecole.]
[00:11:56 The Government of Canada wordmark appears.]
