The Governance Problem

WEBVTT
00:00:00.080 --> 00:00:08.800
Hello, I'm Stephanie, and I'm here again with Erik to talk about the latest and greatest in the localization and AI space.
00:00:08.800 --> 00:00:13.119
So today we're going to be talking about something called governance.
00:00:13.119 --> 00:00:18.719
Now that's a big word, and we're talking about it specifically in kind of this AI localization field.
00:00:18.719 --> 00:00:22.399
But let's go ahead and set some groundwork before we jump into that.
00:00:22.399 --> 00:00:32.240
Eric, AI translation has improved dramatically with large language models, are now often outperforming the traditional machine translation.
00:00:32.240 --> 00:00:37.359
And now they can also be customized a lot in ways that are really incredible, I think.
00:00:37.359 --> 00:00:45.520
But you have been saying in the past that companies are still struggling to operationalize AI in localization.
00:00:45.520 --> 00:00:50.719
So if the models are getting better, what's actually going wrong here?
00:00:51.439 --> 00:00:54.960
There's plenty of discussion about how the models are improving.
00:00:54.960 --> 00:01:00.079
And we have good data that we use to measure that improvement.
00:01:00.079 --> 00:01:07.680
So I'm not going to go into kind of edit distance reports or kind of quantification of how these systems are improving.
00:01:07.680 --> 00:01:14.959
It is clear that they are improving, but the base models themselves still have limitations to them.
00:01:14.959 --> 00:01:22.719
And I think that the LLMs give a lot more control, things like tone and audience adaptation.
00:01:22.719 --> 00:01:32.159
There's more terminology control, there's domain-specific prompts, but they're still probabilistic systems and they're still going to have a certain error rate.
00:01:32.159 --> 00:01:33.519
They're not deterministic.
00:01:33.519 --> 00:01:36.879
And just because they did it once doesn't mean they'll always do it.
00:01:36.879 --> 00:01:42.640
There can be semantic drift, the meaning changes, like how the model is interpreting meaning over time.
00:01:42.640 --> 00:01:53.120
And obviously, famously we understand things like bias, hallucinations, incorrect product definitions, rewriting things in new and unexpected ways.
00:01:53.120 --> 00:01:56.159
So all this stuff doesn't really surprise anybody.
00:01:56.159 --> 00:01:59.120
We've been talking a lot about that, and you can talk about it.
00:01:59.120 --> 00:02:07.359
But what we're seeing is how does the system react to changes in this workflow?
00:02:07.359 --> 00:02:15.039
So the type of errors are changing into like believable mistakes, like things that if you read them, it's like that's a perfectly grammatically correct sentence.
00:02:15.039 --> 00:02:17.199
But how do we govern that?
00:02:17.199 --> 00:02:28.159
How do we govern our human review process or our measurement ecosystem and our risk profiles around the use of these models?
00:02:28.159 --> 00:02:37.520
And so I think in general, even though they're getting better, I think we're underinvesting in attention to the governance of these models.
00:02:37.520 --> 00:02:39.439
That's my starting thesis here.
00:02:39.840 --> 00:02:40.159
Okay.
00:02:40.159 --> 00:02:45.680
So when you say governance, what does that mean in like the basic sense?
00:02:45.680 --> 00:02:51.759
Is this just talking about workflow policies, or are we talking about some more technical things?
00:02:51.759 --> 00:02:53.759
What's the definition of governance in this case?
00:02:54.080 --> 00:02:59.599
Definitely has nothing to do with parliament and it has nothing to do with democracy, nothing to do with current events.
00:02:59.599 --> 00:03:02.159
But governance is actually a complex and interesting field.
00:03:02.159 --> 00:03:04.800
There's a lot of different layers to it, but I'll talk about three of them right now.
00:03:04.800 --> 00:03:06.719
One of them, which is model selection.
00:03:06.719 --> 00:03:09.280
So, for example, different models behave differently.
00:03:09.280 --> 00:03:14.319
There's latency and quality, trade-off, they perform differently depending on the domain.
00:03:14.319 --> 00:03:18.240
So there's some elements there of just which model do you use for which application?
00:03:18.240 --> 00:03:19.759
Just decisions, right?
00:03:19.759 --> 00:03:20.800
Control points.
00:03:20.800 --> 00:03:22.560
Second is model grounding.
00:03:22.560 --> 00:03:29.120
And I'm using this term kind of to address a wide range of things that you can do to help the model perform better.
00:03:29.120 --> 00:03:37.520
It's not just prompt engineering, it's also having to do with supplying the model with the correct terminology authority, product knowledge, regulatory context.
00:03:37.520 --> 00:03:41.840
These are all things that a lot of LLMs can build into their product ecosystem.
00:03:41.840 --> 00:03:45.840
But there's a place where you can put this stuff and the model will perform better.
00:03:45.840 --> 00:04:00.800
So think RAG, think terminology databases, think knowledge graphs, think product documentation, all these things can be used to inform a stronger LLM output, sometimes astonishingly good if those root resources are in fact valid.
00:04:00.800 --> 00:04:05.199
So note to self, those things need to be the correct resources.
00:04:05.199 --> 00:04:08.319
They have to be coherent and they have to be relevant.
00:04:08.319 --> 00:04:13.599
So there's governance in which resources do you supply to which projects becomes an important part of this.
00:04:13.599 --> 00:04:21.199
So the model interacts with enterprise systems and it retrieves authoritative data and then it verifies that information before generating text.
00:04:21.199 --> 00:04:22.000
That's the way it works.
00:04:22.000 --> 00:04:25.439
So the best AI systems are not standalone models.
00:04:25.439 --> 00:04:28.879
Most people are just like send it to ChatGPT and get it back again or whatever.
00:04:28.879 --> 00:04:34.639
It's not the way it ideally should work, especially with corporate high-risk content.
00:04:34.639 --> 00:04:40.000
These systems are connected, and governance is the framework by which you put all this stuff together.
00:04:40.000 --> 00:04:44.079
And that gets to my third point, which is risk-based workflow governance.
00:04:44.079 --> 00:04:52.560
So not every piece of content needs to be reviewed the same way, but you need to keep track of that, which ones went through each of these tasks.
00:04:52.560 --> 00:05:01.279
And when you have, say, TMs that you created with NMT or with a lightweight review, you don't want to leverage those TMs into a high-risk content.
00:05:01.279 --> 00:05:04.000
So you just keep track of your different tiers.
00:05:04.000 --> 00:05:13.199
But I think the smartest deployments out there are taking content tiering into consideration, each of which is tagged with a certain risk profile.
00:05:13.199 --> 00:05:20.160
And then you select and balance the cost-benefit of selective human oversight into each of those.
00:05:20.160 --> 00:05:30.240
And again, the governance is all about keeping track of all this stuff and controlling which content goes through which track and who's talking about it.
00:05:30.240 --> 00:05:33.839
And there's a whole bunch of different aspects of this, but those are really the main ones.
00:05:33.839 --> 00:05:44.720
So what you're really trying to do in terms of creating value, and we're going to get to that in a second, but you don't want to over-review low-risk content and you don't want to under-review high-risk content.
00:05:44.959 --> 00:05:47.040
So, yeah, let's talk about value then.
00:05:47.040 --> 00:05:51.279
So I guess you can look at value in a couple of different ways.
00:05:51.279 --> 00:05:53.360
What are we talking about in this case?
00:05:53.360 --> 00:06:00.959
And how is that value coming out specifically in the this orchestration process versus say AI, LLMs?
00:06:00.959 --> 00:06:03.519
We're using those because they're cheaper, they're faster.
00:06:03.519 --> 00:06:04.959
So that's the sense of value.
00:06:04.959 --> 00:06:07.839
But how is this other side of value coming out here?
00:06:08.160 --> 00:06:11.120
So let's first look into what orchestration actually is.
00:06:11.120 --> 00:06:13.600
Which model do you use?
00:06:13.600 --> 00:06:16.959
What information are you asking it to retrieve?
00:06:16.959 --> 00:06:19.519
When should humans intervene?
00:06:19.519 --> 00:06:23.600
And to some extent also, which humans should intervene?
00:06:23.600 --> 00:06:28.240
And that kind of is relevant to our conversation about subject matter expertise.
00:06:28.240 --> 00:06:32.000
And who is the experience that you're paying for reviewing this?
00:06:32.000 --> 00:06:33.360
And what are you tasking them with?
00:06:33.360 --> 00:06:34.800
What control do you give them?
00:06:34.800 --> 00:06:36.560
And so that's also part of it.
00:06:36.560 --> 00:06:43.920
But also the quality signals, like how are you going to measure what a failure is and how and get an idea of how much it costs?
00:06:43.920 --> 00:06:47.920
And then also the learning feedback mechanisms, like how do you improve models?
00:06:47.920 --> 00:06:49.040
How do you improve workflows?
00:06:49.040 --> 00:06:52.240
How do you identify when something's gone off and how do you fix it?
00:06:52.240 --> 00:06:56.399
So these are some of areas that an orchestration conversation should have.
00:06:56.399 --> 00:06:59.680
Like what how are you controlling all these different aspects?
00:06:59.680 --> 00:07:04.240
Exception management, failure, rework workflows, change management, all that stuff.
00:07:04.240 --> 00:07:10.399
So as the coordination is really critical, right?
00:07:10.399 --> 00:07:12.000
Why?
00:07:12.000 --> 00:07:14.399
What is the business objective?
00:07:14.399 --> 00:07:16.639
What is the value that this creates?
00:07:16.639 --> 00:07:20.560
I think one is that we're using human attention wisely.
00:07:20.560 --> 00:07:32.720
So if we're having the right people, we're looking at the right content and using their time effectively, it means not wasting it on non-value adding tasks whenever possible.
00:07:32.720 --> 00:07:36.560
Second is you really want to reduce the redundant QA cycles.
00:07:36.560 --> 00:07:39.759
A lot of these are very fragmented, but some are very formal.
00:07:39.759 --> 00:07:43.279
And QA cycles cost time, money, release.
00:07:43.279 --> 00:07:44.000
It's a pain.
00:07:44.000 --> 00:07:50.319
So how do you detect and fix and prevent from re-occurring the type of errors that are showstoppers?
00:07:50.319 --> 00:07:54.319
There's also the centralization of visibility across workflows.
00:07:54.319 --> 00:07:56.879
So imagine, let's say, we have three different workflows.
00:07:56.879 --> 00:07:59.519
We have high risk, moderate risk, and low risk.
00:07:59.519 --> 00:08:08.480
Each of these workflows has their own kind of tracks, their own characteristics, their own control mechanisms, as all these orchestration elements would tell us.
00:08:08.480 --> 00:08:15.040
So you want to know like, how are you doing against different criteria for each of these tracks?
00:08:15.040 --> 00:08:19.040
Like, how many critical areas did you have for the high-risk content?
00:08:19.040 --> 00:08:25.360
That might be your number one KPI is number of critical fails in your high-risk content.
00:08:25.360 --> 00:08:38.639
Whereas your low risk content might be optimized around total cost per word or even total product cost, because now you can generate content that just be like poof, in a few prompts, you've got an entire article.
00:08:38.639 --> 00:08:42.000
So, how well are you delivering to that business value?
00:08:42.000 --> 00:08:55.440
And it and word rate may be irrelevant, but maybe it is just word rate, is the it's just trying to get from two cents a word down to one and a half cents a word through prompt optimization and kind of good token management and things like that.
00:08:55.440 --> 00:09:15.440
So, in general, the this is hopefully giving an idea that there's fundamentally different ways to think about these, and that if you do have a risk-based model for how you route work, then you will come up with different metrics and different orchestration parameters for each of these models.
00:09:15.440 --> 00:09:21.759
And then I guess the last part really to keep in mind is governance is all about creating enforceable processes.
00:09:21.759 --> 00:09:29.360
So if you don't want any terminology failures, you need to have some control mechanisms to prevent and review content.
00:09:29.360 --> 00:09:44.000
Again, remembering we're dealing with a probabilistic system, we can't be 100% sure that a model has introduced some surprising interpretation of a common word with more than one meaning without 100% human review.
00:09:44.000 --> 00:09:49.919
And in some cases, that 100% human review might even be more expensive than just having a human writing it in the first place.
00:09:49.919 --> 00:10:08.559
So it's very important to keep in mind the essential purpose of the content that you're creating and how are you orchestrating all the different systems that are supporting the delivery of value, which is a sum of all those different characteristics, the right quality, the right cost, the right characteristics of what you're looking for.
00:10:08.559 --> 00:10:13.519
Yeah, so I think just thinking here, this is also a big issue for our industry, right?
00:10:13.519 --> 00:10:17.200
We're dealing with cost and margin compression as an industry.
00:10:17.200 --> 00:10:35.919
So the complexity of the orchestration and dealing with that complexity, which I've talked about before as well, is becoming more and more important to solve because we don't have massive amounts of generic word count to buffer optimization around.
00:10:35.919 --> 00:10:42.720
Like the cost of optimization, let's say it costs $100 and it's a $1,000 project.
00:10:42.720 --> 00:10:50.480
Then you're getting your workflows down, all that stuff is 10% of your total project cost, like 10% of the total job cost.
00:10:50.480 --> 00:10:51.440
Whatever it is, yeah.
00:10:51.440 --> 00:11:07.919
But if let's say you cut the translation price in half, now your orchestration overhead is still $10, but you're now 20% of your total cost, or maybe not doing my math right, but basically you're you could double or triple the operational overhead of that orchestration.
00:11:07.919 --> 00:11:13.600
So one of the things that we all have to figure out how to do is how to orchestrate faster and more efficiently.
00:11:13.600 --> 00:11:25.919
And we need tools to be able to optimize these workflows as quickly as possible and be able to customize the parameters by which we measure the success for each of those types of workflows.
00:11:26.159 --> 00:11:26.480
Okay.
00:11:26.480 --> 00:11:27.600
Well, thank you, Erik.
00:11:27.600 --> 00:11:28.399
I think this is good.
00:11:28.399 --> 00:11:39.519
It gives people a list, maybe, of things to check to say, okay, am I looking at this aspect of governance and this and this as they go down the list of trying to get their AI systems up and running?
00:11:39.519 --> 00:11:42.399
Because I know that's a huge pain point for a lot of folks right now.
00:11:42.399 --> 00:11:45.840
The LLM seems great, but now it's this whole second piece.
00:11:45.840 --> 00:11:48.639
So thank you for coming in and sharing your insights.
00:11:48.960 --> 00:11:50.080
Always a pleasure, Steph.
00:11:50.080 --> 00:11:52.080
Thanks very much and have a great rest of your day.