The Data Deluge and the Growing Need for Innovation

Recently the Economist had excellent coverage on the huge growth in data on the internet in quite some detail. The amount of digital information increases tenfold every five years! They actually suggest that this data deluge is equivalent to the industrial revolution in overall impact, and that we are moving to what Craig Mundie, head of research and strategy at Microsoft calls a “data-centered economy”. He suggests we are at the nascent stage and that the infrastructural and business models for this new world are not well understood and are just being formed. He also says that big societal and macroeconomic changes are coming with the data economy. Data are becoming the new raw material of business: an economic input almost on a par with capital and labor. If you are skeptical, look at at Wal-Mart, Amazon, Google, Netflix, EBay, Farecast and see how data collection, organization, leverage and analysis underlies their core business mission and value.

I have written previously about how this data deluge is affecting the professional translation industry and why change is necessary not only in processes and tool technology but also in the whole view of the professional translation business. I have pointed out:
-- About how enterprises are also facing a huge growth in data volume from both customers and internal processes
-- The changing nature of the information required to build loyalty in the global customer base. The continuing evolution from static documentation to dynamic user and community created content that is considered the most valuable content in customer support is one example of how dramatic this shift is. The word of mouth impact on products and companies in social networks is already powerful and will become even more so in future.
-- There is also now evidence that crowdsourcing is and will continue to be a force in getting things done in the translation world, not so much because it is cheaper, but more often because it increases customer engagement and allows global companies to address long tail issues in a cost effective way. It does not work everywhere, but it is a model worth understanding and using. Also, we are likely to see more groups of highly motivated amateurs focus on large translation projects like the Yeeyan, Global Voices, Meedan initiatives already show.

However, in 2010 in the professional translation world (and elsewhere), people have gotten used to the tools and processes that got us here. If you look back ten years, I would say we’re not doing things too differently from the way we did in 2000. We use essentially the same software and processes we did back then, though things have sped up a bit and maybe TMS systems are taken more seriously of late. It is very much a TEP world that is optimized for the global business and localization reality of 2000. Professional translation services firms try and build value around project management capabilities and un-definable notions of “quality”. It is ironic that an industry “leader” is named SDL, which originally stood for Software and Documentation Localization. Do they really do any more than that today? Part of the website?

So how would one contrast what happens in the data economy with the old way? As the Economist points out:

Google applies this principle of recursively learning from the data to many of its services, including the humble spell-check, for which it used a pioneering method that produced perhaps the world’s best spell-checker in almost every language. Microsoft says it spent several million dollars over 20 years to develop a robust spell-checker for its word-processing program. But Google got its raw material free: its program is based on all the misspellings that users type into a search window and then “correct” by clicking on the right result. With almost 3 billion queries a day, those results soon mount up.

So as we head into this data economy what will professional translation companies need to do to thrive? It is clear that this is a time for innovation and real fundamental change, in both process and focus. A recent LISA report on Crowdsourcing has this statement from a senior Dell executive in the executive summary that summarizes what global enterprise buyers are looking for.
Dell: “What we want eventually from our services provider is a combination of localization, machine translation and crowdsourcing services.”

So what will a next generation LSP that thrives in the data-centered economy look like? Here my thoughts on this:
-- Competence in managing crowds and ensuring translation quality and managing the overall contribution and validation process 
-- Competence with machine translation, especially SMT which is a data-driven approach that will soon overshadow the RbMT approaches
-- Competence with recruiting, assessing and managing both professional and amateur crowds to engage in translation projects when and where needed
-- Competence with building and managing motivated and engaged online communities and social networks
-- Competence in building and managing large linguistic data repositories that can be brought to bear for different client needs
-- Large linguistic corpus analysis, cleaning and preparation skills
-- Skills in the development of systems infrastructure to enable large groups of professional and amateur translators to collaborate on large, data-rich and very high volume (10M words+)  translation projects
-- Systems process development skills with social networks and crowd initiatives and large data set management skills including handling audio and video translation cost-effectively and efficiently
-- The ability to make rapid and accurate quality assessments on work product in a regular, consistent and definable way
-- The ability to provide a satisfying and mutually beneficial experience for freelancers and amateurs that engage with the firm including useful and free translation productivity tools

This little promotional video for the book Different by Youngme Moon of the Harvard Business School, I thought had a very compelling message that is especially pertinent to the professional translation industry today.

Innovation is something we should all be thinking about and I would bet that collaboration is something that will help and further this exploration. I look forward to sharing this journey with you.

Linking Translation Quality to Business Purpose

The question of translation quality has been a hot mess for the professional translation industry for some time now. It is regularly a subject of discussion, a core theme in every industry conference and yet is also continuously the greatest source of confusion or disagreement amongst translation professionals. This is even further complicated when considering how to determine what an acceptable quality measurement is for many machine translation projects.  One major source of confusion is caused by the conflation of process quality standards with linguistic quality standards.They are different and should not be mixed and matched. The SMT developer community has added to the confusion by introducing automated translation quality measurements like BLEU, METEOR and TERp. I have attempted in this blog to try and clarify some of the SMT stuff at least. But to be able to develop useful MT systems it is necessary to get some rough working definitions in place so that work can be planned, done and fairly compensated. 

Quality Definitions that do NOT seem to work
For the most part, certification in ISO 9001, EN15038, Microsoft QA and LISA QA 3.1 does not provide a buyer with a clear sense of the actual translation quality that will be delivered, even though it does suggest that the certified LSP has process discipline and a quality sensibility. Also, automated measurements like BLEU have very little value to translation professionals as general  linguistic quality measurements and should be used with care anyway. This is why buyers always ask for test/sample translations. They need to have some understanding of what they can expect before they undertake a larger project.

The most widely accepted quality definition that I am aware of in the TEP world is that quality is achieved when the buyer or the customer is satisfied.  Making content useful for a target customer is perhaps a way to approach MT projects as well. Since MT is unlikely to produce the quality produced by a typical human translation process it does not make sense to always use the same quality assessment procedures unless the goal is in fact to get to a human quality level. Wherever high quality is required more human post-editing support will be required, but there are many business use scenarios where lower linguistic quality can adequately accomplish business objectives.  However, in every case we discuss below, the objective is to produce better quality than the free translation services like Google, MSN and Babelfish as I assume that competent professional involvement will produce systems that outperform these free alternatives.


I thought that it would be useful to present three different scenarios where MT could be used, and then see how the business objectives impact the translation quality requirements.

 

eDiscovery & Litigation Support

Business Objective: Identify Key Documents and ensure Searchability of Multilingual Content. Speed & Fast Turnaround Are Critical.
Translation Quality Required: Raw MT produced by a lightly customized engine that has paid particular attention to accurate translation of key terminology and search terms is adequate. Once key documents are identified they go through a more standard human translation process.
Typical Business Use Scenarios: Very large content, very short time frame for turnaround and rapid identification of key documents e.g. Litigation Support and other eDiscovery applications. Also valuable in Military and National Security applications that involve information triage of foreign language blogosphere or websites to “find a smoking gun”.
LSP Role: Translation professionals can play a significant role in developing critical terminology and working on the MT engine to get it to acceptable quality levels for the business application. In patent litigation some TM preparation and training may also be required.

Knowledge Base Support Content

Business Objective: Provide Current and Comprehensive Technical Support and Knowledge Base content to global customers in target languages. Rapid Availability and Breadth of coverage are critical and should be aligned and equivalent to English content. The success and ability of the customer to perform self-service and solve technical problems is critical.
Translation Quality Required: The most active and most critical content is translated by humans. For the bulk of the KB content, raw MT produced by a carefully customized engine that uses localization content, active KB article translations and key terminology as training material is adequate. The content can be linguistically flawed and imperfect but still useful as long as the basic meaning is accurate. Initial content is put into production and made available quickly and can be further refined and improved gradually over time. Users are warned when content is raw MT.In 2010 possibly 1 billion+ people will use this type of translated content.
Typical Business Use Scenarios: Very large content, very short time frame for turnaround and continuing updates on a very regular basis. Major IT firms like Microsoft, Intel, HP have a huge need for this to reduce support costs, improve support experience of global customers and facilitate support  processes for key resellers in global markets. Any industrial engineering / manufacturing company that has a global customer base and wants to improve the technical support experience could be a viable user and many will do so as the success of initial projects is publicized.
LSP Role: Translation professionals can play a significant role in translating most active KB articles, critical content (financial transactions and security related), developing critical terminology and tuning the MT engine to get it to acceptable quality levels for customer acceptance testing. It is possible that they have an ongoing role in improving the MT engine to periodically re-translate the whole corpus as a higher quality level.

High Volume Localization Content

Business Objective: Improve production and efficiency of critical dealer channel service and maintenance documentation for global customers in several target languages. Cost and Quality are as important as Speed to ensure brand identity. OEMs who rely on globally distributed service and maintenance network need to provide regular updates on maintenance documentation to their dealer channel with new models of major products.
Translation Quality Required: Brand consistent human quality is required. Large high quality TMs and glossaries are available to train MT systems. MT engine is carefully developed and tuned to maximize the raw quality and ensure high productivity post-editing effort. MT + Post-editing _ proofing process is used to quickly bring MT output to human quality levels.  Same quality checks are applied as would be to a fully human process and the buyers will have the same expectations.
Typical Business Use Scenarios: Large content (500 to 3000 page manuals), rapid turnaround expected but not real time, and periodic updates as models change. Major Automotive manufacturers, Electronic Equipment OEMs have an ongoing need for this to ensure that their service and dealer channels can properly service new models.
LSP Role: Translation professionals need to play a significant role in building the MT engine and ensuring that it produces high quality output that requires a minimum amount of clean-up work (post-editing).  LSPs need to monitor and validate that they are getting  increasingly higher levels of productivity. This should make each successive version or model easier and more cost effective to produce. It is reasonable to expect new manuals to be done faster and more efficiently.

Thus we have a relatively new concept and approach to quality as the use of MT expands the scope of translation projects. Some refer to this less than human quality  as “usable” and some call it “good enough”. There are many who find this expanding use of MT very threatening, but the greatest use of MT so far is still in areas that just could not be done any other way. As skill levels amongst translation professionals increase and the MT quality improves, it is likely that we will see MT being used in many high volume localization projects, driven primarily by cost and production efficiencies. The industry has yet to really develop clear quality definitions so that buyers and service providers can build effective business models around these new possibilities. However, given the content deluge that we increasingly face, this is coming, and this is a significant opportunity for those service providers who take the time to develop the skills to undertake these and other high-volume high-value kinds of projects.

How Translation Professionals Can Get Started With SMT

At the recent ATA-TCD conference in Scottsdale I noticed that there was a strong focus on translation automation. This was nicely summarized by Ben Sargent of CSA. At the end of my presentation I noticed there were many LSPs who are interested in better understanding machine translation and getting more involved with MT but were not sure how to proceed. Thus, I thought, it might be useful to present a list of resources that might be helpful. I will also raise key questions in affiliated areas, that I think we are all just beginning to explore.

 

It is a clear that I have a bias in favor of statistical MT. This is because I think that SMT provides much more user control, more rapid and continuing MT quality improvements and a greater scope and role for translation professionals to add value and build long-term leverage. My view on the RbMT vs SMT debate is clear, but I would still recommend that potential users explore this on their own and come to their own conclusions.  I believe that the best MT systems we will see in the coming years will have significant LSP/professional translator involvement and that MT will remain a marginal technology for professional use, until it actually makes sense to professional translators. I think that MT systems developed in close collaboration with translation professionals will the best systems around and that they will be domain focused. Quality is key.

So on to some high quality resources: The American Machine Translation Association (AMTA) is reaching out  and trying to better connect to the ATA by timing their annual conference to coincide and expand collaboration with professional translators. AMTA plans to hold several educational sessions tailored to the professional translation audience. I think the quality of the information on business use of MT is likely to be better than typically found at major localization conferences. Even though this conference has been very technical historically, Alon Lavie assures me this is changing.

This is my short list of the best links for web based MT resources:

Automated Language Translation group in LinkedIn where 1600+ professionals discuss many MT related subjects on an ongoing basis. Read through the most active discussions to get a sense for the issues from many different viewpoints. (Requires membership)

Common Sense Advisory provides research on MT in business use for the professional translation industry. LISA also has MT focused content, gathered over the years on how localization professionals interface with MT. The TAUS site also has useful content on enterprise use of MT, though I think that their technology advice is somewhat naive and I do question their editorial policy. (Some content may require membership)

The eMpTy Pages Blog has a lot of SMT related material in bite sized chunks that I will continue to update, and keep focused on why it matters for the professional translation industry. (Blatant self promotion!)

There are also a growing number of video lectures that can be an educational shortcut and way to get informed. Here are a few to start.

For more technical material:

John Hutchins has created a comprehensive archive of MT articles over the years. The Statistical Machine Translation site is also a great source on all things SMT, including links to parallel corpus, open source software tools and research and overviews of the technology.  The Euromatrix project is also a source for useful research, data and findings.

There are also of course some company websites that provide useful information on MT from the perspective of the localization industry. The best content I have seen is at Lionbridge and Asia Online and the Language Studio site which I know will continue to be enhanced with new content on an ongoing basis.

So how do LSPs and professional translators add value when working with SMT? As I mentioned before, with RbMT there is little beyond building dictionaries and human post-editing that a professional can do to improve the linguistic quality of the MT output or the overall engine. Most of the RbMT systems out there today are a result of decades of refinement and work.

In contrast, there are many places where professionals can add value to the "Hybrid SMT" system development process. Some of this is new linguistic work and these new skills will need to be learned and developed. All of the following activities can have a direct and measurable impact on the quality of a hybrid SMT engine:

- Data Cleaning, Preparation and Analysis of Training Corpus
- Development of Test and Tuning Data Sets

- MT Translation Quality Assessment & Evaluation

- Linguistic Analysis focusing on Error Analysis & Correction 

- Dictionary & Glossary Development

- Amplifying Post-Editor Corrections to improve SMT engines

- Ongoing Management of Data Resources &  Linguistic Assets

- Managing Optimization of Domain SMT Engines for a Specific Customer

- Identification and Preparation of High Quality Target Language (Monolingual) data

- Development of Linguistic Rules Structures to improve Quality with Languages like Chinese, Japanese and Arabic e.g. SVO = SOV.

- Creation of Specialized Linguistic Data to Correct Specific Linguistic Error Patterns

 

Remember that the better the MT system is, the higher the raw MT quality will be, and thus the easier and more efficient the post-editing experience is. A good MT engine  provides both a cost and speed advantage to it's users. Most SMT engines can be improved on a regular and continuing basis via the linguistic work and input described above and so it is reasonable to expect ongoing improvements in post-editing efficiency.

These are new skills, and translation professionals should understand that there is long-term leverage in building these engines skillfully. A good MT engine becomes a strategic advantage. It gives you a continuing cost and delivery time advantage in specific domains. And thus we see that Language Service Providers can become Language Solutions Providers (as my friend Bob Donaldson says) who assist their customers in making different types of high-value content multilingual. The future of professional translation has to be where the highest value content is, and increasingly this content is voluminous and volatile. Being able to respond to high volume content dissemination with speed and relatively good quality, becomes increasingly valuable to the global enterprise. Partners who can facilitate this will be valued. The world is changing and user documentation and critical localization work is not going away, but it will increasingly be under cost pressure. The dialogue with the global customer is becoming increasingly important, and in the Web 2.0 world, customers expect and even demand that they get the same information that English speaking customers do. This will not be possible without ever improving and evolving translation automation. This technology provides a major opportunity for translation professionals to build sustainable competitive advantage and differentiation.

The keys to success do not lie with SMT expertise alone. As Jost and I pointed out at the ATA-TCD conference, in addition to translation technology, there will also need to be effective data sharing schemes, improved engagement and management of communities and crowds, and the development of collaborative technology platforms. While many translation professionals view these trends as threats, I am sure a few will realize there is real opportunity here and may even yell out: “There is gold in them thar hills.”

But as we all know gold is hard to find without digging. So start digging.