Hello World Conference: 12 Lessons From an Insightful Event
The Hello World Conference that Phrase organised in Hamburg, Germany, has just finished. After 7 hours of thought-provoking presentations, we feel we have learned a lot.
Representatives from Xing, Deliveroo, AWS, Jimdo, Auto1 Group, Beluga, Lengoo, Wieners&Wieners, Hamburg University, Dublin City University, Otovo, Glossa Group, and Phrase were present and provided incredibly valuable insights into what the localisation process means for them and how they implement it. The talks covered use cases, success stories, business tips, best practices, in-depth research, future challenges, and much more. Keep reading to find out the most valuable lessons we take home with us!
In this post:
- Local is the New Normal
- NMT Is Not Suitable for All Texts
- Good NMT Needs a Lot of Data
- Errors in NMT Are Harder to Detect Than in SMT
- NMT is the Best Choice for Highly Perishable Content
- Anticipating the Need for Localisation Can Make All the Difference
- Everyone at the Hello World Conference Agreed: Localisation Goes Beyond Translation
- In-Company Localisation Needs Centralisation
- Localisation Managers Are Superheroes
- Launching a Product and Solving Issues Along the Way Sometimes Does Actually Work
- Translation Quality is Risk Avoidance
- AI and MT Are Shaping the Future of the Language Industry
1. Local is the New Normal
Davide Gallo, from AWS, shared with us what localisation looks like at Amazon. He made reference, among other things, to the Can’t Read Won’t Buy research by CSA. He also told us that 51 % of people prefer content in their language even if they speak English! Now… how do you effectively localise your content when your catalogue includes billions of products, with thousands of them added by the minute?
Amazon is one of the best examples of why Neural Machine Translation is sometimes the only feasible way of localising copy. Davide emphasised the significant advantage of NMT in terms of speed and cost, and showed us how the output of Amazon Translate is much more natural-sounding than the output of other open MT solutions. We had no idea their software actually ranks in the top 5 % in the MT market!
2. NMT Is Not Suitable for All Texts
Regarding quality and accuracy, Davide explained that their output can be up to 85 % as good as a human translation, although it’s not perfect. We’ll be cheeky here and doubt this percentage, but we agree that the results can be amazing. Davide did express (and he’s right) that companies should only rely on MT for content where they don’t need 100 % precision. That way, it’s best to leave Marketing texts and sensitive information in the hands of a human translator.
Finally, we learned that NMT is useful for a wide range of activities beyond the localisation of product descriptions on the Amazon platform: it can prove useful for other tasks such as media analysis, e-Discovery, content moderation, understanding customer issues at contact centres, text-to-speech conversion, etc.
3. Good NMT Needs a Lot of Data
The Hello World Conference also included research insights from university professors and PhDs. Postdoctoral Researcher Meriem Beloucif, who conducted research on Machine Translation from a Low-Resource Language Perspective, blew our minds. She proposed (and we agree!) that the key element for a translation to be useful lies in the semantics. Preserving meaning when reexpressing the message in the target language is what makes a translation useful.
Did you know that Neural Machine Translation and Statistical Translation systems can be trained? And that, to do that, you inject them with lots of data for the specific language combination and field in question? Jonathan Wuermeling from Lengoo actually suggested that the minimum data input should be between 200k to 400k words.
Imagine the challenge, then, when training systems for low resource languages such as Hausa, Uzbek, Tigrinya, Oromo, etc. How do you improve and evaluate MT output when you don’t have enough data? Meriem Beloucif proposed the first models that successfully inject semantics instead of BLEU while training MT systems. And she showed that, by only injecting the target language (English) semantic bias, translation quality improves significantly.
Ultimately, in Beloucif’s words, what we all want to get from a translation is who did what to whom, for whom, when, where, why, and how?
4. Errors in NMT Are Harder to Detect Than in SMT
Professor Joss Moorkens, from the Dublin City University, brought very comprehensive research on NMT to the Hello World Conference. We analysed the varied ways in which NTM can be put to use and we compared the outputs of Statistical Machine Translation vs. Neural Machine Translation. We also looked at the time post-editors spend at editing each type of output. And here’s the curious finding: errors in NMT are harder to detect than in SMT because the former’s output shows higher adequacy and fluency, and fewer word order and morphological errors.
5. NMT is the Best Choice for Highly Perishable Content
Another valuable lesson from the Hello World Conference is that the “shelf life” of a certain text can determine the choice of human translator vs NMT. Think of a TripAdvisor review and of a website’s landing page headline, for example. Which one is more “perishable”? For which one is high quality more important? NMT is a viable solution for more short-lived content such as a review. The reason is simple: accuracy is not as important for such type of texts as preserving the gist of the source text message. But using the same approach for Marketing copy can be a very, very bad idea! So no, using NMT for your ad copy is not a smart decision.
6. Anticipating the Need for Localisation Can Make All the Difference
Tilman Büttner from Xing spoke about the mistakes product managers should avoid. We actually took a trip down memory lane to when Xing was OpenBC! Analysing the bad decisions the company made back then, he reached important conclusions. We particularly loved this one: “you need to think beyond country borders.” It may seem obvious, yet many brands fail at it.
The most important lesson from Tilman, though, was that developing a product with the potential of reaching foreign markets requires anticipation. Localisation needs to be integrated into the product or feature development timeline from the very beginning. This will help in avoiding half-translated products and an underestimation of customers’ needs in terms of user experience.
7. Everyone at the Hello World Conference Agreed: Localisation Goes Beyond Translation
This might seem pretty basic, but for many, it sadly is not. As many speakers like Tilman Büttner, Anne-Sophie Delafosse (Deliveroo), Antonella Zagaria (Auto1), and Eike-Marie Eiting (Jimdo) pointed out, localisation must consider much more than just the text. Aspects like colours, navigation flow, location of CTA buttons, among many other things, need taking into account when localising content. Even the choice of the product name! We had a good laugh when Antonella showed us the following example:
In many Spanish-speaking countries, “pajero” translates as “wanker”! Imagine how much money Mitsubishi wasted in rebranding the product. And they could have avoided it with some initial research. But hey, lesson learned, right? It reminds us of Honda not changing the name of the Fitta in 2001 when releasing the car in Sweden. In Swedish, “fitta” is a vulgar way of referring to a woman’s genitals (what is it with car manufacturers and sexual connotations? ?) Again: including localisation at the very beginning of the product development timeline is essential.
8. In-Company Localisation Needs Centralisation
Eike-Marie Eiting, Anne-Sophie Delafosse, and Antonella Zagaria brought us a lot of food for thought. The former two are Localisation Managers at Jimdo and Deliveroo respectively, and Antonella is a Product Manager at Auto1, All three of them showed us that localisation can be challenging. Organisations often neglect localisation or handle it in a very decentralised manner, which is a recipe for chaos. Who hasn’t heard the typical “X person in the X department is bilingual, they can translate this”?
Centralisation of all localisation efforts is key to captivate foreign markets and ultimately generate more revenue. Consider Auto1’s operations in 30 European countries and 25 languages: no centralisation would have been a disaster! A common challenge for Eike-Marie and Anne-Sophie was translation requests coming in from all over. They had a lot of tidying up to do when they started at their current roles! It’s interesting to point out that, in both cases, their roles didn’t exist a year ago. Also, all three speakers agreed that centralisation generates reliable data. The benefit of that? Among other things, you can analyse how much the company spends in translation and what is the ROI.
9. Localisation Managers Are Superheroes
The (impressive) work of a Localisation Manager involves the creation and maintenance of style guides, glossaries, translation memories, among other resources. They also act as intermediaries between the teams who need localisation and the linguists who provide it. This way, when someone needs a translation, they will fill out a request and submit it to the Localisation Manager. At that point, this person will assign the task to a linguist or a team of linguists, and will later deliver it to whomever submitted the request.
The benefits of centralising localisation include taking translation off the plate of each department, thus freeing up their time. This way, they can focus on their actual jobs and what they do best, In addition, translation speed and quality improve significantly, which ultimately impacts the product’s performance in the target market.
10. Launching a Product and Solving Issues Along the Way Sometimes Does Actually Work
We know, it’s shocking. But Rikard Eide, a Software Engineer from the Norwegian solar panel start-up Otovo, proved with his experience report on the company’s expansion that this actually worked for them. His presentation took us through the brand’s journey and their expansion from Norway to Sweden and, later, France. Rikard did see, however, that investing in localisation before entering the Spanish market will only result in risk avoidance and higher conversions.
11. Translation Quality is Risk Avoidance
For us linguists, this is not a new concept. But David Benotmane from Glossa Group made this extremely clear during his presentation. After Meriem Beloucif’s definition of quality in terms of semantics, David approached quality in terms of the potential negative consequences of a bad translation. To minimise risks with regards to product liability and safety, and for regulatory compliance purposes, companies need to translate their technical documentation. This is especially important in cases where a flawed translation can result in injury or death.
He went on to describe their QA model and system “myproof”, where the core processes of risk management according to ISO 31000 are applied to processes in translation projects. A risk matrix allows for a risk analysis of the texts that need translating and, consequently, for the development of risk-based processes for translation projects. The implementation of comprehensive risk management for translations results in regulatory compliance as well as a higher quality of translations.
12. AI and MT Are Shaping the Future of the Language Industry
This was the topic of a panel discussion at the end of the conference. Jan Hinrichs from Beluga, Ann Huels from Wieners&Wieners, and Jonathan Wuermeling from Lengoo discussed their own experiences with AI and NMT, both good and bad. It was a great way to conclude the conference as we were all able to realise that, however much improvement is still needed, AI and MT are here to stay and to reshape our industry. In our own opinion, contrary to many linguists who see a threat in these technologies, we see an opportunity: there won’t be fewer jobs for linguists but, rather, the nature of our work is changing and we need to adapt in order to survive.
We’d love to hear your thoughts on Neural Machine Translation! Leave your comments or get in touch with us directly through social media.
Author: Maria Scheibengraf
Maria Scheibengraf is an English-to-Spanish marketing and SEO translator specialised in software (SaaS, martech, fintech), and Operations Manager at Crisol Translation Services, which she co-founded in 2016. With a solid background in programming and marketing, Maria has an in-depth understanding of the technical intricacies involved in software programs, websites, and digital platforms. Maria is also the author of The SEO Translation Bible.