S03E07: Effective Machine Translation Implementation Strategy

Kristin Gutierrez

Thank you, thank you! So good to be here. Thank you so much for having me.

 

Sultan Ghaznawi

I’m so happy you came back to us to speak with us. I’m very excited, but first of all, for people listening for the first time or who haven’t heard your previous interview, can you quickly introduce yourself to them?

 

Kristin Gutierrez

Sure, uhm nice to meet everybody. I am a 16 almost 17 year veteran in the localization industry. Uh, I started off at SDL right after the Trados acquisition. So that’s years ago at this point and I worked my way around the vendor side and the client side. Having worked at NetApp inside their globalization team and I currently serve as vice president of sales within United Language Group.

 

Sultan Ghaznawi

Tell me what are you busy with these days? What is keeping you occupied?

 

Kristin Gutierrez

What’s keeping me busy these days is helping clients try to differentiate what one vendor offers versus another, ultimately which is trying to help them get to their goals. Right? So the clients all have end goals of we need to get more content into more of our customers’ hands in other languages in the languages they prefer and our budgets are shrinking or we just need to do more with the same. Those are things that are keeping me in my team up these days. How do we help them solve those complex problems.

 

Sultan Ghaznawi

It’s no secret in our industry that you’re perceived as an expert in solving problems and not selling solutions. So tell me how do you do that? Because I’m interested to learn what is your trick to learn about the customers’ problems and solve them?

 

Kristin Gutierrez

So for me the a trick to helping clients solve problems is really outlining an outcomes based framework which allows us to together tackle like what are they trying to achieve. Like great, I know I need translation and I know I need to localize the content, but like how and why, and like what’s the right methodology for helping me, my client accomplish that for us, it’s walking them through an outcomes based framework to say OK, let’s run a series of tests, let’s look at different translation methodologies, let’s look at different the language outcomes per you know, the different translation methodologies. And really let’s roll up our sleeves together to get the right quality at the right price with the right outcome. And I feel like that’s usually a combination these days of using machine translation plus human translation in various different capacities.

 

Sultan Ghaznawi

Thank you, Kristen, and that brings us to my next question today we will be covering the subject that has not been covered before yet. All language companies deal with it one way or the other and you just mentioned it. So machine translation or human translation or a combination of both. Let’s talk about what’s an effective way to implement machine translation at a high level.

 

Kristin Gutierrez

Right, at a high level to implement machine translation, you really want to look at curating and augmenting. So the idea is based on existing translation memories that you, the client, have combined with the right methodologies with your vendor partners. How can you train your machine ahead of time, how can you curate that machine to become smarter and that means running the translation memories that you do have through the machine translation application and then augmenting the output on the back end sometimes the output is pure curated raw machine translation, but often it’s combining that machine translation output with human translation. Maybe it’s a translate edit step on top, or maybe it’s a post editing step, so that’s kind of where I see it. Implementation would require some sort of curation and some sort of augmentation combined together to get a win win.

 

Sultan Ghaznawi

Let’s talk about costs or the big elephant in the room. Everyone thinks that machine translation equals immediate drop in cost. Is that true? How do you quantify that to clients?

 

Kristin Gutierrez

I think the… I think it depends on the volume that you currently are translating and what your quality expectations are. So we are really great at walking clients through a loaded cost per word model. So you might be paying $0.20 a language for X, Y or Z plus project management plus different service levels on top of that, engineering or desktop publishing, but you might actually have a weighted word cost of $0.05 a word, and that’s because creatively with your agreement and intent, the technologies behind the scenes being used are translation memory and various levels of human translation. So the cost impact upfront might be minimal if your volume isn’t extremely large, but for certain over the course of a year, if you’ve previously been doing everything using human translation, you will see a cost reduction utilizing machine translation even if you’re doing machine translation plus translate edit or MTPE in the background, you will inevitably see cost reductions it just might not be immediate and so large upfront if your volume is not that large.

 

Sultan Ghaznawi

So with machine translation there is some training involved, just like you’d train a person to do a job. The machine has to learn that too, so there is some sort of investment required to get it up and running. Sticking to the subject of cost, would the cost drop as you move along? The machine becomes smarter, right? It has to learn about your content.

 

Kristin Gutierrez

That is a good question, yes, so if you’re retraining your engine, so if you have a lot of output that you’ve curated your engine and now you’re using the engine, and if you’re retraining the engine every six months or more frequently, depending on the volume that engine is getting smarter over time, so the output is getting smarter and then it can be start to be applied in other departments across your organization, so within one division within your organization, you might see cost savings of X, but now being able to apply the machine translation to different departments across the organization, you might exponentially see cost savings. Prohibitively where you were not being able to do translations before now, you can leverage the machine and do a translate edit on top to still get great quality because it’s trained with the other divisions, translation memories, but the cost isn’t as much.

 

Sultan Ghaznawi

Kristin, let’s talk about how translation companies or the supply side implement MT today. So language or translation companies are still seeing machine translation as a black box. Some of them have no idea what goes on under the hood. So tell me how they can benefit from implementing machine translation into their workflow effectively.

 

Kristin Gutierrez

I think it’s in every LSP’s best interest to take a look at machine translation to implement it, but I also think it’s in the clients’ best interest to be transparent in how you’re implementing it and where. So ultimately, if the question is, how can other LSP’s implement machine translation, then you’re looking at what’s available on the interwebs, right? So you’ve got machines built by Google and Microsoft and other huge platforms. We all want to be able to leverage those machines. And you simply as an organization, as an LSP, want to be able to have like you ULG, we have data scientists, we have SMEs in our Barcelona office, who are are very technical and are very linguistically tuned into all things machine translation. So find yourself a SME who can plug into available networks of machine translation and really get deep in with all the learning there is around machine translation to figure out the best practice to then apply that to your own organization, because I think it will help you be more efficient as an LSP, but then also your clients will gain efficiencies as you’re rolling out, you know, new innovation for them.

 

Sultan Ghaznawi

Absolutely, I couldn’t agree with you more on that, but there are two things as you just mentioned, there is the business side of machine translation, which is the efficiencies that you gain. But then there is also the technical side, which is completely a different beast. There is data science and algorithms to be developed, so not all machine translation engines or models are created equal if you look under the hood, there are different algorithms and then you have different technologies such as transformers and neural machine translation and so forth. What should translation managers know about machine translation to decide on an effective implementation?

 

Kristin Gutierrez

The key here is to not be afraid of looking at it. It truly can be applied to all content types now. Content types where in 2005 you would say, well, no no MT is really great for techdocs, but it’s not good at all for marketing content. Now machine translation there’s a use case for it, with live chat with training you see a lot of augmented AI voice actually even happening now too so it’s not just written, but it’s also spoken, but ultimately I would say take a serious look at what machine translation can do for you and your company to help you realize your goals. Is it cost savings or is it doing more with the same? Or is it doing more with less? And then what are the strategies that you could deploy alongside machine translation to help you realize those goals. It could to start of it really could must be uh, curating your machine and then applying an augmented strategy of PE on the back end, or what JLG likes to do is, uh, translate edit on top of that machine.

 

Sultan Ghaznawi

Sticking to that point, Kristin, you just mentioned that at the beginning it was considered that machine translation may work very well for technical content, but not marketing. Are there certain languages or types of content that MT may not be a good option for?

 

Kristin Gutierrez

I’m sure my team will say yes all day long something, but I would really argue that it depends the engines have gotten so good because of the Googles and the Microsofts out there doing. Other thing that I would say take a look on a client by subject matter basis. It’s really easy to run things through a test you don’t have to curate the engine ahead of time to run it through a test, but run it through a test to see what the outcomes are. We like to do pass or fail and if they pass, at what level are they passing? Are they you know good to go or are they just quasi like decent? And your output might not need to be 100% human, and it might be OK for that language combination in that subject matter. But I would truly say take a look at the subject matter, the client industry, and the language and run some tests in order to see if you should exclude languages.

 

What we like, just to conclude this, is we don’t have a one size fits all, so a lot of our clients now that we’re rolling out all of this innovation, we have it’s kind of like choose your own adventure, but like choose your own translation methodology. Let the data speak for itself, so this subject matter and this language pair equals this type of translation methodology. And you, the client, don’t have to worry about figuring all that out. We’ll align that all in a matrix and you’ll get the quality you’re expecting at a per language level, even if some are doing purely human, and then others are doing a combination of machine.

 

Sultan Ghaznawi

I think we all agree that MT presents lots of efficiency gains. What bottlenecks exist today to deliver on that promise? Do you see, for example, human output capacity, availability of certain languages or type of algorithms as a bottleneck that may be restricting the realization of those efficiencies?

 

Kristin Gutierrez

Truthfully, I see an opportunity. I, I really think that and it’s not, you know, in in all of it I see there’s an opportunity we’re doing things right now in as an industry level that we’ve never been allowed to do because the machines weren’t there yet. The human, you know, there’s so many more human translators now, I truly think that the any bottlenecks we are running into are simply more opportunity to create different workflows or tweak continue to tweak and refine the technology to make it better, but we shouldn’t look at the bottlenecks as like a reason to not deploy a machine translation solution. We should say doing it this way, even if it’s a small margin of content that we’re able to pass through our machine. It’s still, that’s still 5% more than we could do in 2015, right? Or it’s 50% more than we could do in 2005, so I really think that there’s an opportunity at hand and as an industry we are leading this whole idea of how implementing AI can transform the way we’re doing business.

 

Sultan Ghaznawi

Going back to implementation, who should be responsible for training the engine maintenance and retraining within a language provider organization?

 

Kristin Gutierrez

Want to make sure that you partner with a vendor or you know, language services provider that has the skilled resources like ULG does, we have the subject matter experts. We have an entire office in Barcelona dedicated to it. So ultimately, your language partner should be responsible for training your engine, but that ultimately as from a client perspective uhm, gathering translation memories that you’ve done with other translation agencies and allowing one vendor to sort of centralized and manage all of those assets, even if ultimately work is still being funneled to other vendors, having the translation memories every month or quarter or every six months centralized in one place and then using those assets to curate the engine, style guides and glossary’s and translation memories. Then the translation partner is able to tweak and, then the translation partner is able to refine and hone the curation of that engine. So really I believe it’s on the translation partner to do that.

 

Sultan Ghaznawi

And in order to implement a proper machine translation strategy, what type of planning should you be thinking about? For example, how many people do you need to start with a specific language to train the engine, to overtime perform quality control, to make sure the output meets human consumption level quality?

 

Kristin Gutierrez

I think we should get one of my subject matter experts from Barcelona to talk to have like an MT 2.0 with you because I think a lot of the questions that people might be wondering from a from a client perspective. If you’re a buyer looking for machine translation, you need to have satisfaction in knowing that your vendor absolutely can handle everything from A-Z, right? But if you’re looking if you’re a client or a vendor who’s really looking under the hood to say how is that car running? What’s making that car go? Then I think really like a data scientist from my team is more of a subject matter expert who could talk about what the team dynamic looks like and how many people are doing the things we have team leads and I work with one or 2 of our team leads who then manage this whole team in the background. Whenever I have a new initiative with a client, I say listen like I think that we’re going to test out different various levels of machine translation and translation methodologies. So let’s take a look at what that might look like and then my team goes off and does it, but I don’t know, one, two five, ten to fifteen people that would work for you versus that would work for us.

 

Sultan Ghaznawi

So that’s a really good point Kristin. To drive it a bit more home for an organization that doesn’t resources or skillset to implement an effective or proper machine translation system of program, would it make sense to abstract that by working with an organization like ULG that has all the skill sets and resources and capabilities behind the scenes, but all the customer sees is the content going in and translated content coming out without worrying about what goes on under the hood?

 

Kristin Gutierrez

Right, I think there’s a lot of white glove service happening in the industry in general, companies creating technologies that other companies white label to use and resell without, you know, without any funny business, it’s just there’s a lot of collaboration in our industry. So yes, I think that there’s an opportunity for collaboration on this, for example, even within our interpretation line of business we personally, we service other very large language service providers, like very large vendors give United Language Group their overflow for interpretation because they can’t handle the volume or the language combination or whatever the case is. So from a interpretation, some of our clients aren’t just end clients, some of our clients or other translation vendors. I could see the same thing happening with machine translation, for example because it is a very specialized service and certain you want to be able to go to that you want to be able to walk your client down a path to say you know, even if it’s not me who can answer all of your questions, I have a resource of team behind the scenes who can support, and I don’t think it matters where that team comes from, right?

 

Sultan Ghaznawi

So staying on the same topic, what about ready-made MT engines? I mean, none of those exist outside the industry. These solutions or engines are there, but how do you curate your own MT engines leveraging from both generic training as well as your own custom?

 

Kristin Gutierrez

Some companies who sell machine translation, if maybe that’s all they do, they say hey we have access to 30 different databases, so there’s generalized vocabulary from a machine translation that’s out available on the market and we or other vendors leverage that generalized vocabulary. The magic happens when you curate the engine, right? It’s really taking. Let’s say you’ve worked with a client for a couple years and cumulatively they’ve spent a million dollars or let’s say every year they spend two million dollars a year with you. That’s a lot of translation memories to be able to build a curated machine, uhm, so the idea of curating an engine it, you don’t have to have translated millions of dollars in order to have translation memories that are worth curating you can curate engines on as little as 20,000 segments, or 50,000 segments, and even if you’re not there yet with that many segments per language you can still put a plan in place to build out a process for when the curation will start to take on, knowing that the raw MT output, MT that’s not leveraging any commercially available dictionaries and glossaries and terminology etc., that the output might need to be more heavily edited on the front end. But as the segments increasing and increase we can look to curate the engine and the curation of that engine is your company specific content. It’s leveraging both the industry-wide like the Googles and the Microsofts of the world and then also your specific content. So it’s both, right?

 

Sultan Ghaznawi

Kristin, machine translation by itself has come a long way, and certain models have demonstrated impressive results on the blue score is an end to end solution within reach for mid to high complexity type of content and stand alone, MT without post editing?

 

Kristin Gutierrez

Absolutely, absolutely I see because we have a methodology in which we really follow an autcomes based approach. You know, I don’t see a client who has never done translations before looking at machine translation as a catchall for them, I really see the solution that you’re proposing can there be content that’s smart enough that gets put out there that doesn’t require post edit? Though that’s really for clients who are doing pretty big volumes of work. So they’re spending a million to five million dollars a year on translations, then from their perspective, they have various levels of quality requirements they have they have white glove quality for like high, you know it could be press releases or it could be certain training materials , or it could be like letters from the CEO that absolutely require that full touch, but there could be a whole level on the bottom, support articles, knowledge based content, et cetera that ultimately it’s better to just get content in your buyers hands in their designated language then it is to make sure that it’s dotted all your I’s and crossed all your T’s.

 

And that in between there is everything that’s getting edited from the output, right? So you’ve got full human at the top, at the bottom You’ve got the raw output and then in between you’ve got machine translation that’s been post edited or translate edited on top of so absolutely I think if there is a organization with a large enough volume of content that would be good enough quote UN quote. It’s like what’s under the tip of the iceberg if all the content underneath the tip of the iceberg would benefit from some sort of language, yes, I think the engines are already there.

 

Sultan Ghaznawi

So today I see that raw machine translation output is used in certain low exposure or low criticality type of content such as Facebook posts or certain emails. Where do you see some of the typical applications where our clients can drive value from using the raw MT output and their enterprises and their organizations?

 

Kristin Gutierrez

Great question. Of course it depends, but I think you hit on some of that. I think it’s knowledge based articles, I do think it’s live chat, I think it’s chat logs. I think it’s technical reports. From there. I think it depends, and interestingly, I think in five, ten or fifteen years from now it’ll be a whole different answer, because we are today with the answer to this question where we were in 2005 was like can we apply machine translation to anything. Back then we were barely confident as an industry to be able to apply MT to Techdocs. Now we’re saying everything under the sun could benefit from some combination of a machine, but you’re asking what can benefit just from raw output, or hopefully, if it’s an organization, again, that’s translating in other parts of their business, the output, technically, is never raw. Because once you curate the engines for your organization you’re able to leverage those. Yes, does training speak differently than marketing differently than HR differently than tech docs all day long… so there might be segmented machine translation engines within the umbrella of a machine, but could you turn that machine that’s curated? Can you turn that big machine translation database on for knowledge based articles for live cat for chat logs? Yes, and would it be way more catered? It’s it better already than raw because it’s your content.

Sponsor

This podcast is made possible with sponsorship from Hybrid Lynx, a human in the loop, provider of translation and data collection services for healthcare, education, legal and government sectors. Visit HybridLynx.com to learn more.

 

Sultan Ghaznawi

Speaking of editing the output of a machine translation, MT by itself is not sufficient. More than often you need to clean up the output for human consumption, or in other words, post edit it. Tell me why post editing machine translation is different from editing professional translation created by a human being?

 

Kristin Gutierrez

I think it’s just ultimately, ’cause they’re looking for different things with translate edit as a process as a two step or translate edit and proof as a three step translation process you are expecting you, the translator and you, the client, are expecting a certain level of quality. But when you, the client knows that it’s a machine translation, either curated or not, plus a post edit the output is also meant to be different, so every time you step down or step up, it depends on how you look at it but like every time you change the translation methodology your quality expectations are supposed to change. It’s not for better or for worse.

 

It’s just that certain content requires 60% quality whereas other contents you can’t let out the door for less than 99% quality and so that content would require human translations. But there’s a threshold in there that you, the client, need to decide what is my threshold is it 90% eighty percent 75%? And then what is the output? And as a result, the editor who’s doing the post editing? They’re not trying to change the machine translation output to be human quality, they’re not trying to change it to be 100% quality, they’re simply trying to edit it to get to be 75% or 85% quality. If that makes sense, so they’re looking for they’re looking for different things. So if they start to over edit, right? Now and and like you’re not happy you’re in country reviewers aren’t happy like you’ve done a machine translation and you’ve done a post edit and it’s and the in country reviewers are still pushing back. And then you edit it again and you edit it again and you edit it again. Now you need to just take a step back and say hold on, is this methodology work, is it right for this type of content? ’cause it probably isn’t.

 

Sultan Ghaznawi

Kristin, and on the same note, let’s say you ULG has a client that is buying raw machine translation output from ULG and they have their own people who will edit or correct this output. Does a quality convention exists between ULG and it’s client to agree as to what is acceptable by the clients so that ULG can either improve or constantly meet that target? What does that quality model looks like and who defines it?

 

Kristin Gutierrez

Well, are you alluded to this earlier in that there is an industry scoring called TER [Transaltion Edit Rate] score. The translate… translation, edit distance and then the BLEU scores. So there’s a combination of statistical data That you can look at from the machine translation output That again, my sneak and talk to my data scientist team can talk to but you Want to look at the TER score and the BLEU score to create, what we’ve simply done in certain clients is do yellow or green, yellow and red. Green means it passes all day and you look at the data, it’s not a subjective pass or fail, it’s objective… it’s literally just based on data. And then if it’s yellow, it means there might have been linguistical errors or terminology challenges with the machine and yata yata yata, and then red means it fails and even if it fails it might mean the engine needs to be trained or retrained to your point, so I would always encourage clients to not like starting with a raw machine translation output might not get you the places that you’re trying to go right away, but if you can look at curating your engines and applying an intelligent process to like analyzing the output from that. So you’ve curated the engines and then you objectively look at the output using tear score, Bleu scores, and like everything in between, now you can start to say OK, this is the best fit for my content or my corporation.

 

Sultan Ghaznawi

So on that note, let’s say you ULG obviously has to clean up their output from machine translation to a client who is looking for a final quality translation. ULG will have to use post editors. What makes a good post editor? And in the past hour linguists have been editing human produced translations, right? Should translators relearn how they work with organic versus artificially translated text?

 

Kristin Gutierrez

I think there’s a ginormous opportunity here also. So at ULG we’re focused as a group and as a strategy, we think that we’re looking at the tipping point of our industry and how it changes the roles of editors and linguists across the board. So the the idea is that the the primary concern I think originally from an MT perspective, is that it would reduce the value of the translator, or the linguist or the editor, right? And then ultimately they would as you’re saying, we need to temper their expectations or change their skill set. But what if the clients became more intelligent about how a curated machine could truly benefit brand engagement. What if machine translation could entice that organization to do more multilingual SEO, what if it meant they could do more localization? What if it meant that… and especially if you’re a client that’s been doing translations from humans for several years and your content changes and evolves over time, right?

 

But like ultimately like, let’s say you have robust translation memories and you throw that into curating machines for different languages, etc. let’s say you’re able to just get machine translation content out there and then the linguists and your language partners and their editors and everybody in the supply chain, now they’re focused on augmenting the output. Now they’re really focused on true localization because we think as an organization as the tipping point kind of evolves, the editors won’t have to relearn how to edit or how to temper the quality. Actually, they’ll get to paid, potentially paid more to help clients really tap into that customer journey globally and to really like tackle things like SEO and like a real localized form because it’s less about…. at that point it becomes less about OK quality, expectations or 100% for human and like 75% for whatever MTPE and you know 50% for raw MT. What if they level set? This is interesting, right? There’s I think there’s an opportunity there for the editors and for everybody in the industry.

 

Sultan Ghaznawi

Speaking of which, while we are still talking about post editing, do you think our industry needs a convention or set of standards to ensure output quality and naturalness across different types of texts, different languages, and so forth? Because it is a completely new area in terms of how we transform a mechanically produced content to human consumption quality level.

 

Kristin Gutierrez

Yes and no. My sales brain wants to say no. Because there shouldn’t necessarily be a standard because standards should be derived by the client’s intent and they should rely on their language partner to be able to help them narrate that path, but over the years, as I’ve grown more of an operational side as well, I think yes, like standards always make everybody happy. Everybody wants a little bit of structure. And it is kind of, it is a unique perspective to think OK whether I’m doing post editing with one vendor or another the output I would expect would be similar, but I truly believe my sales brain takes over again and I truly believe that like the client, not that the client’s gonna tell us, like they’re they’re asking us, to say what’s your experience in X? Well, how should we feel about Y and we as the vendors have a responsibility to say hey, we’ve seen this we’ve been there. We’ve done that, this is what worked for everybody else but I bring it back, let’s test it and let’s see what works for you and like the editor should… I mean, I think just like the translator is right at the end of the day, like the supply chain, there are metrics in place. There are standards in place. There are expectations.

 

Sultan Ghaznawi

So for folks on the client side, how do they use machine translation engines that are trained by their LSP partner for their next use cases? Let’s say it is in healthcare and ULG develops this model for this specific client. Do they access these models or engines using API’s? Does it make sense to host their own models trained by ULG? What is the standard access model?

 

Kristin Gutierrez

It it can be both. It can be one. It can be either. I actually have clients who, I don’t want to give away specifics, but yeah, they uh, it’s a combination of, right? Ultimately, ULG differentiates by truly… I mean, we’re we’re very agile so if the client wants to connect via API, we have that ability. If they want to buy something and use it on their own, we have that ability. If they want a managed service, this that’s kind of the sweet spot of a lot of the vendors I’ve worked at in the past. It’s really like let us work together. We’ll host a lot of things on our end like we try to take the heavy lift off of our clients to say like you do your job and let us do ours, which is the whole managing of the lifecycle of translations from A to Z and how to access the machines, but we absolutely also sell the licenses so that they can self access because you kind of hit on it earlier, raw or in their case curated machine via a license that could… they sometimes you don’t have time to ask your vendor for what the answer is because you needed the answer within the second so they can absolutely for us log into a platform and get instantaneous machine translations on the fly for any content type for any language. And if we haven’t curated the language for them, it’ll give them more of a raw output. But if we have curated their engines, they’d be able to tap into that so it wouldn’t even be raw at that point. But other clients, like I said, we they have connected through API and then others. We just manage A to Z.

 

Sultan Ghaznawi

On the same note, if they decide to self-host a model or even access through API’s, what is the process of retraining? Who takes that responsibility and does that client learn to do that by themselves, or does it make sense to them to go to their vendor like ULG and ask them to please retrain because we have this much more data coming in now?

 

Kristin Gutierrez

I mean then there there are clients out there who are super savvy and knowledgeable in all things globalization and manage the technology and do a lot of things themselves, and a lot of those vendors come from the operational side of client, of vendors or they’ve worked their way up on the operational side of the client within globalization. But I really feel that those are, it’s a minority of who’s using this, and so ultimately it’s a big, it’s a big to do to set up a globalization department to set up people and train people and hire the right people. And even when you do have all those people and I’m thinking back in my personal client side experience. There was an in, it was like a ten-person team that managed globalization but they still weren’t responsible for training their own engines, even though at that point they were already using machine translation their vendor and vendors were still responsible for training the engine. So I think it’s even if you have a lot of skilled resources on the client side internally in your department all working together, you’re ultimately on the client side you’re working toward KPIs and you’re working toward pushing this big boulder up a hill, which is proving to your C-Level that localization isn’t a cost, that it’s really driving value to your organization, and how are you doing that? You’re helping the company increase sales globally, you’re helping to, you know, the ROI or you’re tapping into the global customer journey. So your role I truly believe from my past on the client side is to communicate that to your C-level. And then you rely on all of your partners like your translation partner to do the heavy lift and the heavy lift would include curating and managing and retraining the engines. I don’t, I don’t think they would want to do that.

 

Sultan Ghaznawi

So if you were to do retraining regardless of whether you are on the client side or supply side, how often would you update an MT engine?

 

Kristin Gutierrez

Well, that’s a debatable topic. Uhm, if I generically said every six months, I would say most people would agree with me. But I’ll say there’s absolutely outliers to that where the cadence, or the type of content, or the reality of what you’re doing requires a much more frequent duration, but I wouldn’t say it’s every day because it is a heavy lift and that one would be logical every week, every month, but look… but most often it’s every six months.

 

Sultan Ghaznawi

Kristin, you alluded to this earlier, but I’m going to ask you this question anyway. Do people need to worry about the technology architecture behind the machine translation model? Every vendor today claims they use the technology stack of X or Y and so forth. Should I care if it is a neural network based model or not? How how should I spend to learn about that tech stack or architecture behind these solutions?

 

Kristin Gutierrez

If I’m a client, I would want to spend my time making sure that the quality outputs are aligned with what my objectives are for that particular content type. Uhm, and do I want to know something about the translators themselves? Do I want to know something about the machines? Do I want to know stuff about the technology? Yes, all day long, right? I probably have certain SLAs I need to meet internally or I’m just might be a curious buyer for example. But ultimately, a company like ULG, who does have this huge team we have a lot of teams that are really tiny groups that are subject matter experts within themselves, and that’s how we’re able to like present this white glove boutique service level even though we’re a top 20 LSP. So we can answer and we would love for a client to ask those nitty gritty questions because we can absolutely back it up. But there there are vendors ultimately that might you know, not have all the resources internally, and it might be hard for a sales person like me to articulate all of the answers, if the client is asking if I don’t have a subject matter behind the scenes who I can pull into the discussion so, so yes, it’s fair for the client to ask. But I think it depends on what really they’re trying to achieve by asking.

 

Sultan Ghaznawi

Do clients really need to know whether a translated text was created by machine translation or a human translator? If they only want the text to be correct, accurate, error free and contextual right doesn’t make a difference how the translation was obtained?

 

Kristin Gutierrez

I mean ultimately the architecture or what’s behind the hood does have an impact to the scale and performance of the system. Uh, meaning it can affect the way the engines train the accessibility data retention, the control, the speed and these are all critical factors in making sure like what you, the client signed up for is what you’re expecting. Uhm, so I would say yes, I would say yes, clients should be made aware. I would say yes from a client perspective, when considering your goals for MT, there should be consideration into like what goes on behind the scenes, making sure that your partner, your vendor partner, does know what they’re doing. But ultimately I do think it’s a case by case basis.

 

Sultan Ghaznawi

That brings us to the end of this interview. Kristen, my last question is to ask for your advice and future direction of this technology. Where do you see it heading?

 

Kristin Gutierrez

Well, I I do think that it’s absolutely here to stay in our industry and it’s very exciting for me to have seen the evolution of it, at least since 2005 to now up. So I would say get with your vendor and have them roll up their sleeves to look at in its simplistic form, like what is a translation methodology map that you could do so that you know the machine translations, the engines can be curated and you’re not spending all of the money now on human translation that you’re really looking at various combinations of humans plus machines plus various outputs of that machine to help you realize your goals without sacrificing quality where it cannot be sacrificed. I just think it’s a really exciting opportunity right now for our industry to just dig in deeper to machine translation and the linguistic, the linguistic side of the business, hopefully it adds value to them, because maybe it allows them to truly get back to their roots. True localization, helping with this CEO, really helping customers communicate more effectively in global markets. So that’s what I’m looking forward to and that’s where I think we’re headed.

 

Sultan Ghaznawi

Kristin, what a great conversation. I really enjoyed talking to you about this fun and exciting topic. I think everyone listening today on both sides of the localization business found action items that they will take with them and apply to their practice. This is what knowledge sharing and industry level education looks like and with that I want to thank you for your time and for contributing to the collective knowledge of our Industry.

 

Kristin Gutierrez

OK, you always put me in the hot seat. Thank you so much Sultan. It was great to see you and hear from you!

 

Sultan Ghaznawi

OK, it’s time for my roundup of the interview and my analysis as to what has been discussed. While the world has caught on to artificial intelligence and machine learning, our industry is lucky to be the first actual beneficiary of AI. In the 1950s when Marvin Minsky and his colleagues were proposing theoretical implementation of artificial intelligence, the main use case was language. Over the years mechanical translation has evolved and machine translation became a commercial offering well over a decade ago. Today it flourishes in many forms and its applications are very vast. Translation providers have been offering both raw and finished or post edited machine translation for a few years now. The efficiency gain both in time and monetary value makes an effective machine translation worth it. The caveats, however, are many, and we can easily get carried away with over training of models, or ignore privacy implications by using customer data for training the model. That said, we will see new and exciting applications of machine translation in the months and years to come.

 

Kristin Gutierrez gave us some interesting ideas and thoughts that we should consider in using automation to our advantage. If you have any suggestions for discussion topics or guests that you would like to hear from, please let us know.

 

Don’t forget to subscribe to the translation Company Talk podcast on Apple Podcasts, ITunes, Google Podcasts, Spotify or your platform of choice. Give us a thumbs up or five star rating wherever you’re listening!

 

Until next time!

 

Outro

Thank you for listening.

Make sure to subscribe, and stay tuned for our next episode!