In my last blog Much talk about Machine Translation, I suggested that the best approach to knowledge base (KB) translations, or any bulk content translations, is to divide, prioritize and conquer.
Human translation costs typically correlate to the number of words needing translation. A quick way to cut down costs without sacrificing quality is to reduce the word count to translate.
A knowledge base is often divided into two parts: the problem definition part, sometimes known as the questions, followed by the solution part, or answers. To find a solution to the problem in a KB, one must first find the problem. Once located, a solution can be viewed and studied.
A typical ratio of the number of words to translate relating to the problems vs. the solutions (questions vs. answers) is roughly 1/3. In other words, 25% of the word count in the KB belongs to the problem definition part and 75% to the solution part.
A user that is entering a query to the KB wants to do so in his or her native language. To stand a chance to find a match, the translation of the problem needs not only be accurate, but 100% consistent with the text used in the product (i.e. GUI and error messages used by the software). This is why we recommend that the problem definition part (25% of the word count) is translated by human translators (HT).
For the solution part (or answers), we recommend applying the Pareto Rule. If 20% of your KB entries are queried 80% of the time, applying human translations to the 20% that is most used will result in accurate responses to 80% of the queries performed by your users. In others words, if you are to apply statistical analysis anywhere, it is early in the process where it can have the best impact!
Using this process, translating 40% of the KB content by HT will correctly answer 80% of queries, while resulting in immediate 60% savings in translation costs.
For the remaining 60% of the text, you can consider optimized Machine Translation (MT). But proceed cautiously. MT on its own may be successful in giving the user a way, at times humorous and at times frustrating, but a way to perhaps get the gist of the meaning.
Before MT is to be used, you have to optimize it by including all your software and language glossaries and translation memory (TM) and make the user decide if and when to use it. A policy of never forcing MT on end-users and never serving it the same way you serve your quality source and quality translated text will reduce your liabilities caused by errors in the MT output and will maintain your investment in your company’s and product’s image and brand.
So this is the process we recommend:
- Clean up your KB source and give your users an easy online KB interface.
- Allow users to view the source language if they wish to. Many international users, particularly users of scientific, engineering or other B2B applications, even with limited knowledge of English, will understand English better than the machine translated output.
- Require users to perform the MT task themselves. Since this is an automatic process, don’t translate and store, translate on demand. This way you are not publishing the inaccurate translations and users will use it only at their own risk. This will limit your liability. Also, if MT algorithms improve in time, the output will improve.
- Provide your users the ability to request professional translations if the machine translation output is incomprehensible. A submit-for-translation button linked to your support or translation group will allow submitting a ticket and prioritizing the translation process.
- Give your users the ability to suggest corrections or improvements to the MT translations via a wiki. Submitted translations are later moderated by your support or translation group and published, increasing the accurate translations content. This is a way to implement crowdsourcing inexpensively and successfully.
Your users should be provided an environment that is conducive to quality improvements over time.
This will demonstrate your dedication to quality, service and your commitment to their long-term interest.
The intent of your multilingual KB solution is to educate your international users on how to best use your product and to self-solve their problems. So track that usefulness and compare the reduction of support calls as the quality of the KB content translations improve. A properly translated KB will tremendously reduce your support costs where they are most expensive– overseas.
Do this and you may come to the same conclusion Aristotle came to millenniums ago: The roots of education are bitter, but the fruit is sweet. Yes human translations are costly, but their dividends well justify their costs.