Translation and Localization powered by STAR Transit

Watch the short video for more information on Transit functionality and usage: youtu.be/D7_pJCQ7N8s

www.star-group.net

In the past I've repeatedly written about the new kind of technology that Lilt has brought into the translation space. I have also written about how Lilt is an example of a young company with large outside investments that has continuously reinvented itself. Now, on the heels of its largest round of investments yet led by well-known investment company Sequoia Capital, Lilt is showing itself in its latest stage of an ongoing metamorphosis that might well shake up the industry.

Over the last few weeks as they were getting ready to make this latest move public, I had a number of conversations with folks from Lilt to ask why they went this route and why they think this is finally the way to operate successfully.

So, what is it?

Lilt has slowly rolled out a "managed service" division that combines translation services provided by freelance translators with the technology offering at the core of its business. So it's just another technology vendor who has also turned into an agency? At first glance it might look like this, but here is where it's different and where it really could have an impact on the world of translation. Translators in the supported language combinations (which have grown to an astonishing 29 languages and 58 language combinations with English <> Afrikaans, Chinese (Simplified and Traditional), Czech, Danish, Dutch, Finnish, French, German, Greek, Hindi, Hungarian, Indonesian, Italian. Japanese. Korean, Norwegian, Polish, Portuguese, Romanian, Russian, Slovak, Spanish, Swedish, Thai, Turkish, Ukrainian, and Vietnamese) will

• do a small test project in Lilt to get familiarized with the environment and have Lilt evaluate the translation;
• continue to work in the Lilt translation environment for free (if you work with Lilt's managed services);
• be paid by the hour at a price that can be negotiated;
• be paid within three to five days of project completion;
• directly and transparently communicate with the end client through Slack; and
• have the liberty to contract directly with the client (if the client is interested in that).

Say that again? I know, I actually had to verify especially the last point a number of times because it seems so counter-intuitive from what we know about our world. The answer I repeatedly received is this: "Lilt is a technology provider (period)."

While providing access to services is a good way to provide access to the technology that Lilt provides, really only the latter is of true interest to the growth of the company (it also has much higher margins). Is this a successful concept? Time will tell, but recent numbers seem to be supportive of this endeavor. There are about 300 translators working for Lilt right now, and in June 2018 a total of 1.2 million words were translated. There are presently a relatively small number of five employees working in the managed services division (of a total of 30), and there is an attempt to create a real sense of community among translators as exemplified in a series of webinars that will be offered to them (and will not "just" be how to work more efficiently in Lilt).

If you are not familiar with Lilt as a translation environment, it might be helpful to read some of the articles I've written over the last few years of its existence (see here, here, and here). While each describes different stages in the long path to becoming what it is today, they all focus on what makes Lilt different: It's a tool that gives access to a machine translation engine, which at the same time also serves as a quasi-translation memory and termbase, which at the same time learns from your translations as you translate and adjusts its suggestions for any translation even as you work in a single segment. It comes down to this: it's one way of using machine translation without post-editing. Or as Spence Green, Lilt's CEO, says in slightly hyperbolic fashion:

"At Lilt, translators will never be asked to post-edit. We believe that the MT post-editing / rate reduction business model leads to the sweatshopification of language work. That's bad for the world. If you're great at your craft, then please join us."

(Even if you don't like the statement as a whole, you gotta love "sweatshopification.")

You've probably already guessed from the positive tone of this article that I'm excited about all of this. And this is true on a number of levels: no per-word compensation (yeah!), a potential disintermediation via an intermediator, and of course a smarter way of working with machine translation.

It'll be really interesting to see what this will do to our "industry."

Oh, and the language combinations that are most desired right now are English> Finnish and English> Dutch (Belgium). 

PDF Translation for Professionals

Translating PDFs is easier and quicker with TransPDF.

• Available within Memsource and memoQ
• Compatible with all CAT tools.
• Fast log-in for Proz members.

FREE for your next PDF project.  

All file formats, all languages, all target groups, better quality, shorter time-to-market -- one standard solution.

. . . is a product that will not be on the radar of the individual translator anytime soon but might very well be an interesting solution for larger companies or even LSPs who have the need to connect to the many different content management systems they have to handle.

You've read many times before (here and elsewhere) that content located in content management systems is difficult to handle. Why? A) Because every system requires a different kind of access to get to the content and B) what do you with the content once you've extracted it and translated it and then need to bring it back into the content management system?

At this point there are a number of solutions, at least for the most widely used content management systems. Systems like Wordbee, Lingotek, Smartling, and even Memsource come with a number of prefabricated connectors. For very popular CMS systems (I'm aware that "CMS systems" is much like my favorite "BFF forever") like WordPress there are a number of more or less usable plugins that deal with translation. Even a tool like memoQ has a well-configured WordPress connector. Other solutions include some kind of a middleware between the content management system and the translation environment and management system like the one offered by iLangL or the Lionbridge-owned Clay Tablet.

Xillio is in the realm of the latter. It's also a middleware product, but it's not specifically designed for translation purposes, even though this is one of its highlighted use cases. The idea with Xillio is that you as the user (more likely: system administrator) don't really have to know much about the content management system (and I mean that not only in the narrow definition as a "proper" CMS but also online platforms like Google Drive or Dropbox or large Enterprise Management Systems) that you're trying to "talk" to. Instead, all you need to know is how to direct Xillio to access those systems and Xillio will translate (so to speak) those commands into the native commands of the system you're trying to access. So, rather than learning 21 ways of accessing data (which is the number of systems presently supported by Xillio), you have one that works for everything. Once you get the data out (presently in the native format but soon also in XLIFF) you can translate it and then import it back into the system where an additional column for that language is created or however the system wants translated data to be stored.

Of course, it would be helpful if all of this were already integrated into a translation management system -- as indeed is the case with SDL Managed Translation (here is a description of what SDL Managed Translation is), and Xillio is "talking to all of the TMS technology vendors."

Some might ask how all this fits in with the ongoing development of TAPICC, a standard-to-be that, among other things, will accomplish something similar to what Xillio is offering: API (application programming interface) standardization between content and translation management systems. I asked Xillio's Łukasz Rejter about this, and here's what he said:

"How we see it is that it will be very unlikely that the CMS producers implement TAPICC. It takes a long time for CAT/TMS providers to do so in the first place. However, we see a great potential in it and plan to support it. This way, we could be a gateway for all the data coming from the content sources, and all the tools that support TAPICC will be able to quickly integrate."

I'm not sure I share his pessimism about CMS providers not being willing to implement the upcoming standard -- after all, they will benefit from it when they can sell their systems as more easily translatable -- but it is certainly true that commercial projects like Xillio can move faster than a committee-driven effort.

Oh, and now that we're talking about being "commercial," there is, of course, the price point (which brings us right back to why -- for this and other reasons -- this is likely not a product for the individual translator). All 21 connectors are offered with a single fee of 10,000 euro/year (this includes onboarding and support) or 6,000 euro/year without support aside from documentation.

25% off SDL Trados Studio 2019 Freelance + FREE AutoSuggest Creator

We've got a very special offer for you. Get 25% off SDL Trados Studio 2019 Freelance and receive AutoSuggest Creator for free.

If you're looking to upgrade to Studio 2019, you can get 25% off all upgrades.

Don't miss out, these offers end soon! Get your special offer here >>

I'm writing this month's column on board a plane on my way home after speaking at an event that may well prove to be a watershed for interpreting in the 21^st century, at least for conference interpreting -- where I have spent my career (Okay, okay, maybe "watershed" is overstating it, but the meeting was definitely a milestone). About a year ago, I was asked by the Italian Region of the International Association of Conference Interpreters (AIIC) to keynote at a two-day continuing professional development (CPD) event on new technologies for interpreters. My topic -- Remote Interpreting: The Current State of Affairs.

But first, a little background. Remote interpreting has never been an easy subject for conference interpreters to discuss. On the whole, I think it safe to say that we conference interpreters enjoy the lifestyle our profession has traditionally afforded (i.e. international travel and the privilege of being at the shoulder of history as it is being made, among other positive aspects). On the surface, the rise of remote interpreting threatens to take all that away. Many interpreters see remote at the first step towards the "Uberization" of the profession that could turn us into a faceless voice available on demand at a moment's notice, with all the problems and difficulties that it implies. In some cases, those fears are not unfounded. So, it should come as no surprise that conference interpreters have been leery of and even pushed back against the use of remote interpreting, sometimes very forcefully. This has led to an often tense relationship -- and on occasion, outright hostility -- between conference interpreters and the technology providers who are making remote interpreting possible.

That's why this meeting in Rome was so important. AIIC Italy was able to bring 50+ conference interpreters from various countries together with representatives of some of the most recognized cloud-based interpreting platforms on the market today.(The remote interpreting platforms at the event were in alphabetical order: Catalava, Interprefy, KUDO, Verso, and VoiceBoxer.) The first day focused on understanding the larger context in which the growth of remote interpreting is happening (remarks by AIIC President Uros Peterc, AIIC Italy President Stefano Marrone, and a keynote by yours truly), the technical standards required to provide interpreters with what they need to do their job successfully (Klaus Ziegler, chair of the AIIC Technical and Health Committee), and other technologies that are helping interpreters become more efficient and entrepreneurial (Marzia Sebastiani). Keep an eye out for a recording of the live webstream from day one on the AIIC Italy YouTube page in the coming weeks.

Day two, which was not streamed live, included demos from the technology providers that were present and, most importantly, an open and honest discussion among conference interpreters and the technology providers. Interpreting between Italian and English was provided on site both days, and remotely during some of the platform demonstrations. Some of us were relying on the interpreting to participate actively in the meeting, as we do not understand Italian, so a big "thank you" to the interpreters is in order. As you might expect, not everyone saw eye to eye, but that wasn't expected or desirable, really.

While experiencing the technology first hand was great, what made the meeting so valuable was the discussion that ensued after the platform demonstrations on day two. Interpreters expressed both their concerns and their excitement surrounding the challenges and opportunities that remote interpreting presents. And the platform providers listened. They really listened and then explained their own hopes and concerns as well.

Here are some of my key takeaways from this event:

• Distance interpreting is extremely diverse in its implementation. If we don't have a clear understanding of the kind of remote interpreting we are referring to, we can end up talking right past one another. Remote interpreting for on-site meetings is what we most often think of, since it has the most immediate effect by changing our traditional work model, but there is a growing number of webinars and web meetings that require remote simultaneous interpretation (RSI). These events often last an hour or less and are a growing source of work. We need to work with the RSI platform providers to craft business models that take interpreters' requirements into account. Here are two that we identified. First, fair remuneration that also recognizes the need for preparation time for these shorter events. And the need for interpreters to show flexibility on how their services are priced to take advantage of this growing market. Traditional day rates for these new one-hour meetings simply won't work.
• The cloud-based remote interpreting platforms are seeing the most growth in the private sector. Interpreting in the public sector shows a slower adoption curve with this delivery mode being employed under very specific circumstances and usually with significant use of hardware, such as multiple video screens and hardwired video and audio connections. These conditions are excellent but extremely difficult to replicate in the fast-moving private sector.
• There was a lot of interest in the "interpreting hub" concept -- centralized studios located around the world where interpreters can go to provide their services at a distance for different kinds of remote assignments. Many colleagues recognized that they simply do not have the appropriate conditions at home or the desire to set up a home office with the infrastructure required for this kind of interpreting. That may change in the future, but interpreting hubs seem to be the way most interpreters want to go. The interpreting hub model also addressed the widely-expressed concern to have booth mates physically in the same place.

Perhaps most importantly, the event organized by AIIC Italy opened up a channel of frank and respectful communication between interpreters and the technology providers. Building and maintaining trust between these two groups will be crucial to the success of these new interpreting delivery models, because the interpreting delivery platforms are useless without the interpreters, and the interpreters will need these delivery platforms if they hope to offer their services on these new markets. In short, we need each other if we are going to succeed in including professional conference interpreting into the evolving communication models of the 21^st century.

Do you have a question about a specific technology? Or would you like to learn more about a specific interpreting platform, interpreter console or supporting technology? Send us an email at inquiry@interpretamerica.com.

The Lenguas 2019 program is now published and registration opens on October 24^th! Choose from dozens of continuing education workshops for translators and interpreters. Register soon, as space is limited. Join your colleagues in Mexico City on January 24-26, 2019 for this groundbreaking conference!

Click here for more information.

Serge Petelo, the co-founder and Business Development Manager of Nucleus, sees his translation environment tool as particularly suitable for (smaller) LSPs -- like Acceant, the company that Nucleus originated from (and they eventually hope to become independent of) -- and "for translators who want to grow into LSPs."

He told me that Acceant built the tool for their own translation and translation management purposes a couple of years ago because they needed to "simplify their review process," their "vendors complained about high pricing of the tools they were using" (SDL Trados, memoQ, and Fluency), and Acceant found it "frustrating to deal with floating licenses" that some of these technology vendors offer (sounds like a demanding client to me...). As Acceant build their own tool, they didn't look at the ones mentioned above but more at the newer generation of web-based SaaS (Software-as-a-Service) tools like Memsource and Smartcat.

Today Nucleus still needs some additional implementations (the termbase really is only a glossary, there are no dropdown list-based searches, and you can connect only one machine translation engine at a time), and there are other features they want to implement (such as accounting services as part of the tool), but I thought the tool was very intuitive and super-easy to use otherwise.

The translation memories (you can use as many as you want at the same time) are written in MongoDB; translation files are stored in XLIFF (which can be exported to be translated externally and then re-imported); configurations and re-configurations of projects can be made at any time on the fly (including the easily configured and then automated workflow steps); there is unlimited archiving of all files (all on Amazon's AWS servers -- both in Europe and the US); every change is recorded and can be rolled back; and the search feature for files, projects, and clients (which I complained has no dropdown list feature above) is actually really cool because it uses a nice and fast auto-complete system.

Unfortunately, that auto-complete system is not available in the translation editor to automatically call up content (where it might be most helpful). When I pointed that out, Serge mentioned that they have not found a way yet to implement that without significant speed losses to a system that is otherwise super responsive.

The interface is almost spartan, which I really like. In the translation editor (the typical grid with source on the left, target on the right), the only controls shown are the ones for the currently selected segment, and even those are very sparse. It's a role-based system, and if you have the appropriate role you can edit TM or termbase matches right and immediately in the lookup fields, which is nice. Another use of the role feature is for the translation client who sees the identical interface that the project manager or translator also looks at, with only projects being displayed that belong to the client.

The list of supported file types is very large, including image-based files that are internally processed with the OmniPage OCR processor (just like Wordfast Anywhere and Smartcat). It's frustrating that access to the whole range of file formats is available only in the corporate editions -- which simply does not make sense. I asked them to re-evaluate that policy, so we'll see whether they will. The relatively high price point for the professional translator edition ($30/month) also doesn't make sense. I also advised them to look at that again, and I think they will for their exhibit at this week's ATA conference. Let's see whether it will stick.

Let's also see whether the tool in general will stick. There is pretty fierce competition in the browser-based translation environment market right now -- but that's good, right? It should mean that the tools are sharpening each other more and more in the competition.

And if you clicked on Nucleus's hyperlink above and now think the team is located in South Georgia and the South Sandwich Islands (which the .gs extension in their URL might easily point to), you're wrong. It's a company in Provo, Utah, made up of graduates of Brigham Young University.

Imagine all the world's best dictionaries at your fingertips!

For a fixed monthly fee, on all your devices, integrated in your daily work applications!

WordFinder Unlimited -- One service, 5 applications and more than 300 dictionaries in 26 languages. WordFinder Unlimited was voted the winner of the Process Innovation Challenge at LocWorld in Warsaw 2018.

Click here to find out 10 things you need to know about WordFinder Unlimited

Meet the WordFinder team in booth #55 at the ATA in New Orleans.

(Note: A version of this article will also be published in the ATA Chronicle)

When using CAT systems in a typical professional translation workflow, several aspects could be improved. The SCATE project investigated several of these aspects, both from a human-computer interaction point of view and from a purely technological side. A comprehensive extended overview of the project can be found in Vandeghinste et al. (2017), or on this website.

Analysis of human-machine interaction and practices of professional translators

The results of an international survey (128 participants) and the observations of translators'/terminologists' work practices (16 participants) revealed shortcomings in the current translation environment tools related to ergonomics and usability, integration of technologies, and performance. Based on the results, we designed an online translation editor prototype to demonstrate how certain aspects could be improved.

Its user interface (see the image below) displays translation suggestions coming from different resources in an intelligible and interactive way. The suggestions originate from different web services, such as fuzzy matching in a translation memory (TM), machine translation (MT), and automatic terminology extraction. MT and TM are combined by selecting parts of a fuzzy match and applying MT on the remaining sentence parts ("pre-translation"). A lookup mechanism highlights terms in the source segment that are available with their translation equivalents in the bilingual glossary. The prototype adds context to clarify why certain suggestions are made. Details about the prototype are available in Coppers et al. (2018).

Intellingo UI

Improved fuzzy matching

Given a sentence to be translated, Translation Memory Systems (TMS) look in a TM for source language sentences that are identical (exact matches) or highly similar (fuzzy matches), and suggest the translation of the matching sentence. Fuzzy matching techniques mainly consider sentences as sequences of words and use very limited linguistic knowledge. We studied the use of syntactic information for detecting TM sentences, which are not only similar when comparing words but also when comparing the syntactic information (such as lemmas, part-of-speech tags, or parse trees). We investigated whether such abstract, syntax-based matching is able to assess the usefulness of matches in a better way than methods purely based on sequences of words. Our experiments show that this is clearly the case. We get even more improvement when sentences are compared in multiple ways, combining different metrics. For technical details, see Vanallemeersch & Vandeghinste (2015).

MT-TM integration

We investigated ways in which TM retrieval is combined with MT by integrating the results of both. The M3TRA prototype consists of two components: (a) fuzzy match repair, i.e., the automatic editing of close matches found in the TM; and (b) pre-translation (see above), in which the translation of parts of a fuzzy match is automatically detected in the TM.

We performed experiments on ten language pairs (English <> German / French / Hungarian / Dutch / Polish) that involve multiple language families. Using a test document, we compared the output of M3TRA to the output of a pure MT system and that of a pure TM: we evaluated the outputs by automatically comparing them to the human translation of the test document (using the BLEU, METEOR, and TER metrics). Significantly higher BLEU scores for nine of the ten language pairs were reported, whereas METEOR and TER scores showed comparable patterns. Full technical details can be found in Bulté et al. (2018).

Informative Quality Estimation (QE)

In order to build an informative QE system, we first annotated a corpus of machine translations for different types of errors, resulting in an error taxonomy. In a next step, translators post-edit this corpus, allowing us to relate specific error types to specific post-editing time. With this information, we develop informative QE systems that predict the time needed to post-edit segments based on the predicted error types. This allows filtering out MT suggestions that are likely to be problematic. Besides predicting the type of translation errors, these systems show where these errors are located in the suggested machine translations. See Tezcan et al. (2018) for full details.

Extracting terminology from comparable corpora

When working on a particular domain, it is often the case that available translated data are too sparse to extract appropriate terminology automatically. We therefore studied how this extraction could be performed using comparable corpora, texts whose source and target languages are not translations of each other but approximately deal with the same subject. We tested a model in which the topic or meaning of the words in both texts is represented in a so-called semantic vector space. Whenever these source and target meanings are close to each other, we can assume that they have a similar meaning and can be translations. We evaluated these models using aligned Wikipedia documents (English-Dutch) in the medical domain. From the English side we selected 500 words that were manually translated into Dutch, and we compared the translations to the automatic suggestions. The approach that yielded the best performance is advancing the state of the art in this topic.

In these data, we saw that morphology is important, so we also combined the above word-based representations with character representations (that's right, representing the meaning of individual letters in context). The combination of word and character representations serves as input to a deep learning network, which outputs a score between 0 and 1. The higher the score, the more confident the model is that the two given words are each other's translations. Our experiments show that this approach outperforms more handcrafted approaches toward morphology. We extended the system so it can also work with phrases (multi-word terms) instead of mere single words. Details are given in Heyman et al. (2017).

Adapted speech recognition as post-editing interface

How can we improve speech recognition accuracy using the source language text and the translation probabilities used in the MT model? In past research, speech recognizers were set up to output multiple possible outputs, and these outputs were ranked using MT probabilities. However, this slows down the recognition, and information may be lost as not all possible outputs can be produced. Instead, we integrated the source language text and MT probabilities directly into the language model of the speech recognizer. This was done separately for each sentence/paragraph. This method reduces recognition errors by ca. 5% absolute and 20% relative on spoken Dutch translations from English. The method assumes that translation consists solely of one-to-one alignments, i.e., each word in the source language can correspond to only one word in the target language text. This is obviously not the case in reality. In MT, "phrase-based translation models" address this issue.

We have integrated phrase-based models into our implementation and extended the recognizer with "named entity models." Named entity models attempt to improve recognition for proper nouns by estimating their pronunciation and language behavior. Many named entities remain unchanged during English-to-Dutch translation, leading to reliable estimates for relevant named entities based on the source language text. Experiments show that the combination of phrase-based translation models and named entity models further reduces the recognition error to ca. 6.5% absolute and 25% relative on the same spoken Dutch translations from English. All technicalities are described in Pelemans et al. (2016).

Conclusions

The SCATE project investigated several aspects of the professional translation workflow and advanced the state of the art for a number of aspects. We have presented the results at numerous academic and industry conferences as well as international organizations. We hope that the CAT tool providers will implement some of the techniques we developed in SCATE so that translators can benefit from them in their daily work.

The SCATE project ran from 2014 to 2018 and was funded by the Flemish Agency for Innovation (IWT-SBO-130041).

Biographies

Iulianna van der Lek-Ciudin is a pre-doctoral researcher and lecturer at the University of Leuven. Both her research and teaching focus on translation environment tools and their impact on the translation process. In addition, she is optimizing and developing new modules on translation technologies and training programs for students and professional translators. In SCATE, she investigated translators' and terminologists' work practices through surveys, interviews, and observations at their workplaces.

Dr. Vincent Vandeghinste is a researcher and guest professor at the University of Leuven and the Dutch Language Institute. He has a master's degree in Psychology and a PhD in Linguistics (Machine Translation). He teaches courses on Machine Translation, Computational Linguistics, Language Engineering Applications, and Natural Language Processing. He was the coordinator of the SCATE project.  

Dr. Tom Vanallemeersch holds a master's degree in Translation and a PhD in the domain of MT at the University of Leuven. Before joining the SCATE project, he worked for L&H (text-to-speech department), Systran (MT development, Paris), and the MT@EC team of the DG Translation (Luxembourg). In SCATE, he worked on the combination of TM, MT, and speech recognition. He recently joined CrossLang (Ghent), a company providing consultancy and customization services for MT.

Register with the super early bird rate before October 31st and save $100

If you would like to promote this journal by placing a link on your website, I will in turn mention your website in a future edition of the Tool Box Journal. Just paste the code you find here into the HTML code of your webpage, and the little icon that is displayed on that page with a link to my website will be displayed.

If you are subscribed to this journal with more than one email address, it would be great if you could unsubscribe redundant addresses through the links Constant Contact offers below.

Should you be interested in reprinting one of the articles in this journal for promotional purposes, please contact me for information about pricing.