SAP’s Translation Solutions: When to Use What
Miriam Kaeshammer (left) and Beatrice Pasch
In times of growing user expectations, a personalized IT experience for shopping, working, and daily life is a standard requirement. It follows, then, that offering a user interface (UI) in the local language is a prerequisite for success in regional markets. While the task of making business global now seems to be easier than ever, integrated translation solutions are having to catch up – but they are doing so, slowly but surely. According to the Globalization & Localization Association (GALA), “the language industry is big business, with the worldwide language services market growing at an annual rate of 5.52%.” Endpoint co-founder Paul Veness quotes research conducted at the California State University that details the widespread acknowledgement amongst global companies of the need to expand into foreign markets, with “74% of multinational enterprises believing it is most important to achieve increased revenues from global operations”. This research also finds that “56.2% of consumers say the ability to obtain information in their own language is more important than price.” In the context of business software, language availability is therefore key to sealing the deal.
In response to this trend, SAP offers two approaches to integrating translation into business solutions: A machine translation API that is part of the SAP Leonardo Machine Learning Foundation, and the SAP Translation Hub. Many customers and interested followers have asked which translation solution on SAP Cloud Platform is the best for their specific use cases. And that’s a fair question. It’s time to shed light on the use cases for both translation options that are currently available at SAP, and also to allow for a glimpse behind the scenes. We spoke to two members of the SAP Language Platform team in SAP Globalization Services (GS): Miriam Kaeshammer, a machine translation expert, and Beatrice Pasch, Product Manager SAP Translation Hub.
SAP Translation Hub versus SAP Leonardo Machine Translation
Question: Bea, as a product manager, you have been approached by customers asking why they should invest in SAP Translation Hub when machine translation, or “MT”, can also be used within SAP Leonardo. What do you recommend them to do?
Bea: That’s true. There are more and more cases in business applications where simple machine translation is sufficient. I’ll give you an example. An English-speaking support employee wants to get a better understanding of a problem a Spanish customer has reported. Thanks to integrated machine translation, he does not need to ask the Spanish customer to enter the support ticket in English but can use the gist translation provided by the MT engine and start to work on a solution. That’s because machine translation provides enough information to indicate what the issue is.
There are other cases, however, where a superficial and quick translation is not enough, such as in the health sector. This is an area where the requirements regarding translation quality are significantly higher. Here, machine translations need to be checked and post-edited by language experts. In its translation workflow, SAP Translation Hub offers a user interface (UI) that colleagues in the country locations and translators at agencies can simply open on the web to review translations online. Excel files no longer need to be sent and shared, and users save time and money because changes can be implemented directly within the web UI with the help of a quality index and status management. What’s more, the SAP Translation Hub orchestration offers customers the option of using their own company-specific terminology and of integrating additional machine translation providers in the future. A lot more features from customer wish lists are already scheduled for release on the SAP Translation Hub Road Map. One of my personal highlights is the opportunity to review translation proposals directly in the context and environment of the application.
Quantum Leap in SAP Machine Translation
Question: Miriam, your team supports both translation offerings. Can you give us some background on how this evolved?
Miriam: My team within SAP Globalization Services develops machine translation services for different applications and usage scenarios. The SAP MT system is based on open-source software components. It implements the wider community’s best practices, and leverages the vast volumes of human translations created over 40 years of SAP’s history to build its engines.
Thanks to a shift in technology, there has been a quantum leap in machine translation quality in recent times: neural machine translation (NMT) outperforms statistical machine translation (SMT) in a wide range of set-ups and language pairs. With NMT, translations are more fluent and human-like than with SMT. Well-known MT providers have been at the forefront of conducting NMT research, developing NMT solutions and adopting the new technology. Since 2017, all our SAP MT developments have been focused on NMT. Currently, SAP MT is available in the following two applications:
- Within the SAP Translation Hub, NMT translations are provided for 88 language pairs: English from and into 39 SAP languages as well as a number of language combinations with German as the source language.
- 18 language pairs (English from and into Chinese, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish) are part of the SAP Leonardo Machine Learning Foundation’s machine translation API.
Other Machine Translation Use Cases at SAP
Question: Does SAP use MT scenarios internally?
Miriam: Yes, the quality of SAP MT has been evaluated in various scenarios. In the area of Knowledge Transfer & Education and openSAP, there have been several promising experiments with subtitles translated using SAP MT (see graphic below). Human evaluation of SMT versus NMT translations revealed a 20-25% quality improvement from a consumer point of view. In addition, a productivity study comparing the throughput (words/hour) of translators post-editing SMT and NMT translations clearly showed that NMT requires fewer corrections and is quicker to post-edit.
Furthermore, machine translation of SAP Notes and SAP Knowledge Base Articles (into Japanese and Portuguese using SAP MT) has been judged to be good enough to offer MT directly to customers as a self-service in the SAP One Support Launchpad.
Question: Can you explain, in simple terms, how you benefit from machine learning?
Miriam: Let me answer your question in the context of machine translation. Machine translation is one of the oldest applications of artificial intelligence, and it is also considered one of the most difficult. In the early days, machine translation centered around the use of linguistic rules: lexical rules, like “the cat” in English is rendered in German as “die Katze”, or “der Katze”, depending on the case required in the given context; and structural transfer rules, like rearranging the English word order of subject-verb-object to subject-object-verb for Japanese. Those rules were manually created by (computational) linguists and translators. Rule-based MT systems provided very accurate translations, but it is not hard to imagine that they were extremely time-consuming, expensive to build, and difficult to maintain and extend.
Statistical Machine Translation (SMT) was an idea born in the 1990s that translation could become an application of machine learning. This idea revolutionized the field. Put simply, machine learning provides a way for computers to learn translation rules from large collections of texts that have already been translated, meaning that linguists and translators no longer have to sit down and write thousands of rules. Various approaches have been developed over the years, all of which come under the heading of SMT. In simple terms, you could say that SMT works by breaking a sentence down into small sequences, translating each piece, and then stitching them together again, like a jigsaw puzzle. Overall, data has become the key to building machine translation systems, and machine learning algorithms the key to detecting co-occurrence patterns in this data.
Neural Machine Translation (NMT): In the past few years, several developments (including more effective learning algorithms, new model types, and better hardware) have led to the successful application of deep artificial neural networks to machine learning tasks, including machine translation. Instead of learning concrete translation rules and statistics about how they are used, NMT utilizes a neural network to model translation between languages. In simple terms, NMT takes the entire sentence to be translated as one unit, generates an abstract mathematical representation for it, and produces a new sentence from there. This new technique has led to a quantum leap in machine translation quality; translations are a lot more fluent and, thus, more human-like.
But let’s remember that, despite all the technological advances, machine translation is not a solved problem just yet. Extensive research and development is required, both within and outside SAP, so there is still a lot of work for us to do.
Machine Learning in SAP Translation Hub
Question: So, machine translation is a task that, nowadays, is tackled best with machine learning. Which other use cases of machine learning emerge within the SAP Translation Hub?
Miriam: My team’s focus is obviously machine translation. However, team colleagues are also working on other topics that involve machine learning to make the SAP Translation Hub more intelligent. Three of these are quality estimation, text classification, and smart retrieval of existing translation options.
SAP Translation Hub or MT API within SAP Leonardo Machine Learning Foundation?
The following overview provides helpful advice on when to use which approach:
SAP Leonardo Machine Learning Foundation – Machine Translation API
SAP Translation Hub
Stay tuned: Subscribe to the SAP Translation Hub Newsletter!