TAUS - Enabling better automated translation

Customer support (CS) used to be all about running call centers. It then shifted to user searches of knowledge bases, with all the complexity this means for multilingual delivery pipelines. And now it has spread to sharing information on social media. Two years ago in Berlin, TAUS launched a conversation with the CS community by joining forces with the IT industry’s Consortium for Service innovation to explore a shared interest in localizing customer support. At the recent TAUS User Conference in Portland, attendees had an update on the CSI support model, and learnt from a successful user case deploying an MT solution.

CSI’s Greg Oxton shared a holistic view of the new “content explosion” for customer support. Today’s document stream includes informal content from many authors. And instead of being marketed top down, brands are being redefined as the conversation a company has with its customers. This means that the one-on-one call center experience is no longer a key function (only 2.4% of CS), and that 97% of the audience now engage with the relevant information via portals and communities on a many-to-many basis. Content is, once again, king.

Surveys show that customer satisfaction is based on customer loyalty (not profitability figures, for example) and that customer loyalty in turn depends on experience of content. The user’s emotional experience of this content is therefore decisive, and the support agenda is a major driver of this experience.

Today, content derived in various forms from the community is three times more important than knowledge found on a portal, which is 10 times greater than call center support volumes. This exponential growth in support content is obviously not uniform in quality. So the challenge facing companies is how to inventory content sources and influence, rather than control, improvements in quality.

One dimension of this challenge is that growth markets where support needs will be greatest are in non-English speaking cultures, which complicates access and understanding of customer needs. Another is that such content types as chat and IM, blogs and wikis are all involved in the content mix, and require tracking, analyzing and evaluating.

Once the content sources have been identified, it is necessary to decide on the right flows. Some content, for example, should be frozen into books and updated at regular intervals; other content will need to propagate more fluidly and unstably across social communities, and any attempt to control this will damage the community. These choices will also influence what should be translated and how: centrally or by the crowd?

One solution is to build a ‘tag cloud’ of all content types and identify the best options for the trajectory of a given piece of support content across the cloud. This kind of tool allows CS managers to bring their content strategy to bear on the lifecycle of their content, and on whether it can be Googled, or translated (or not), so that the form this content takes matches the most important “touch points” in the customer experience as a whole.

Intel presented a working MT solution for localizing customer support content with a 24-hour time lag (instead of the previous 3 weeks) to emerging markets that are uncomfortable with using English language websites. To kick start the process, Spanish, then Portuguese and Chinese geographies were chosen, because they were cultures usually tolerant of MT output, and also because SMT systems worked effectively for these language.

The approach involved taking a “holistic view of the eco-system” and how translation automation plays a role in the create-to-publish chain. The task involved leaving out the translation of sensitive content on warranties and similar, merging translation memories (5 to 10 million words), creating dictionaries, cleaning TMs, delivering the system to the translation vendor for training, installing the trained engine and measuring output using automated and human metrics, and testing user reaction with a satisfaction survey.

The results of various surveys showed that end users in their respective geographies systematically evaluated the output quality as better than the QA reviewers did. Visits to the site rose tenfold over two years, and although there was no side-by-side display of English source and translated text, the sites included a comment box asking if and why the service helped users in their support query. Satisfaction with the service steadily increased over time.

OTHER ARTICLES ON TAUS USER CONFERENCE 2009