As the localization industry adapts to global challenges and changes in the wake of the COVID-19 pandemic, the development of exciting new technologies and the publication of important research findings continue unabated. Join us for a brief look into some of the most important recent developments.
Amazon DeepSubQE: building quality estimation of subtitles
Two researchers from Amazon Prime Video developed and analyzed a new system for estimating the quality of machine-translated subtitles. The study, entitled “DeepSubQE: Quality estimation for subtitle translations,” was published on April 22, 2020. The purpose of the study was to examine ways to improve translation quality through quality estimation. Improving translation quality to boost foreign viewership experiences is an essential component of the growth models of leading streaming services like Amazon Prime and Netflix. The researchers translated some 30,000 English video subtitle files into French, German, Italian, Portuguese and Spanish.
The use of quality estimation (QE), defined as a “machine learning technology that automatically assigns a quality or risk assessment to MT output without having access to a human-generated reference translation,” presents translation challenges. This is largely due to there always being more than one way to translate a sentence.
The results were encouraging. The researchers concluded that the DeepSubQE model was accurate more than 91% of the time. Success rates were even higher for longer sentences. Future success will be dependent on increasing the operational load of the system and the number of output languages.
ESPnet-ST: improving speech-to-speech translation systems
ESPnet-ST is an open source software toolkit that was created to help developers with “the quick development of speech-to-speech translation systems in a single framework.” The software received a significant amount of attention on Github, a software repository for developers that allows the sharing of code to build improved software systems.
A “toolkit” refers to a software library that includes snippets of code that developers can use in their own software programs. ESPnet-ST originally featured automatic speech recognition (ASR) and text-to-speech (TTS) code, but the software was updated to include code for building MT systems. This made it the first toolkit “to include ASR, MT, TTS and ST recipes and models in the same code base.” The researcher-developers are optimistic that ESPnet-ST can help others make the next breakthrough in language-related communication barriers.
Katan Skynet App for Zendesk: switching languages in customer service applications
Brands are placing more emphasis than ever on providing memorable customer experiences through post-sales support. Studies continue to show the devastating economic impacts of providing poor customer service: one report found that some $41 billion is lost by US companies each year following bad customer service experiences.
More and more businesses are relying on technological solutions like chatbots to provide support in real time. An important component of seamless communication in a global economy is the ability to quickly shift between languages. A new app for Zendesk, a popular support, sales and customer engagement platform, developed by Katan Skynet seeks to do just that.
The app gives customer service teams the ability to communicate with clients in their native languages. The customer service agent’s response is machine-translated into the customer’s language, then post-edited by a Zendesk editor. On the customer’s end, he or she only sees the translated version of the response. Likewise, communications from the customer back to the customer service rep will also be translated.
Anycount 4D word count tool: increasing translation profitability
Anycount 4D is a new word count tool designed to help translators increase profitability through accurate word counts. The tool is a standalone program that works on many different file types, including PDFs, text files, images and subtitles.
According to the developers, word count tools for common file formats like Word or Excel are not accurate enough, often missing words and characters that costs translators money. Users can select from monthly, quarterly or annual subscriptions.
Streaming and re-translation: improving online meetings
The Coronavirus has changed the way that the world holds meetings, at least for the immediate future. Virtual meeting platforms like Zoom, LogMeIn and Google Meet have experienced an unprecedented uptick in use. Increased competition in the space has led to the implementation of new features to gain an edge.
Features that providers are working on rolling out include machine translation and automatic transcription to improve the experience of global meeting attendees. A Google study published last month, entitled “Re-translation versus Streaming for Simultaneous Translation,” investigated these two methods of translating an incoming stream of words.
Streaming translation has been the most common method used up to this point. However, its major drawback is that translated text cannot be altered once it has been displayed to users (think closed captioning). The other method, re-translation, offers the possibility of changing text after it has been displayed on-screen.
The researchers analyzed German-to-English and English-to-French language pairs to determine which method offered better quality and latency. Latency, or the delay between the spoken and translated content, is of particular importance to the user experience. The study found that re-translation performed better in both categories.
Re-translation also provided superior quality, since it “produces ‘high final-translation quality’ because the translations are iterated and improved.” Latency was reduced since “it results in less of a delay because it ‘always attempts a translation of the complete source prefix.’” Researchers also found that by limiting the number of changes to the text, re-translation was “as good as or better than state-of-the-art streaming systems.”
It will be interesting to see how providers of virtual meeting services implement new translation features in the coming months as remote work plays an increasingly important role in the post-COVID-19 economy.
As the world continues to assimilate, the need for efficient translation technologies will become increasingly important. Moreover, the COVID-19 pandemic has led to major societal changes creating new technological needs. With more people than ever working remotely and viewing content on their favorite streaming services, these recent advancements in translation technology seem up to the challenge.