Our Analysts' Insights

Neural machine translation has recently garnered the lion's share of attention for artificial intelligence (AI) in the language industry. Highly visible applications such as Google Translate and Microsoft Translator rank up there with self-driving cars in the public imagination. In addition, requirements for multilingual content are growing far more quickly than the number of human linguists, which positions NMT as a needed solution. We also identified a lot of interest in the potential for AI to improve project management and simplify manual tasks such as vendor selection and status tracking.

What gets lost in the well-deserved focus on NMT is that AI is starting to quietly improve other aspects of the industry. In the past few months, CSA Research interviews have revealed different areas where it is making a difference. Here we focus on three companies that are deploying AI-based solutions to address new opportunities where previous solutions have either fallen short or been non-existent.

Memsource Automates Detection of Non-Translatable Content
CEO David Čaněk told us that Memsource has created a team dedicated to building new AI-based features. Its first release is a function to identify segments it recognizes are non-translatable text in almost 300 language pairs. Numbers, names, boilerplate notices, and branding slogans often either remain in the source language or require no translation. Čaněk reports that such text amounts on average to 14% of segments and 4% of total content volume. Although translation memory software has supported non-translatables as a feature for a long time - Memsource's AI-powered feature recognizes more of them, including people and product names, and indicates whether a segment is recognized as a non-translatable (100%) or as a candidate non-translatable (95-99%). Čaněk claims that, thanks to AI, the number of non-translatable matches has approximately doubled when compared to the previous rule-based approach.

Memsource Tags Non-Translatables as 99% or 100% Matches
Source: Memsource
 

CSA Research observes that a similar approach should enable translation tools to handle common - and tedious - changes to otherwise non-translatable content easily. For example, if the system encounters price lists in German with numbers like "4,69€," it could automatically change them to "€4.69" for English-language versions of the price list, thus reducing the need for translators to manually correct such figures. Any developments that free linguists from doing work that machines can do reliably will help make them more cost-effective and increase their value in the process.

TextShuttle Cleans Translation Memory Databases
Zürich-based start-up TextShuttle leverages AI in a totally different way. Over time, translation memories become corrupted with poor and outdated translations that replicate themselves over time and lead to a significant waste in effort and lower translation quality. Samuel Läubli, TextShuttle's chief technology officer, told us his company offers a service that uses machine learning to identify bad TM segments. It does not attempt to repair segments, but can identify them for deletion. TextShuttle focuses primarily on creation of customized MT systems and treats this service as a way to help its customers improve their training data without the need for extensive manual review. Although it has yet to market this as a stand-alone service, it points to the ways in which AI can help LSPs and enterprises address common difficult problems.

Pairaphrase Delivers Assistive Communication Services
The last of these three technologies brings us back to MT, but treats it in an unusual way. Pairaphrase is a new technology offering that targets a market overlooked by most developers of language tools: assisting business users who need to translate as part of their non-translation professional activities. It provides MT results to its adopters and then creates a TM from any corrections or changes they make for future reference. The company thus provides a hybrid of MT and TM for users who do not know or care about the details of translation technologies. It auto-propagates translations in real time based on the corrections the linguist makes, so in this sense it functions much more like an adaptive MT system (such as Lilt or SDL BeGlobal) than either a traditional translation memory or machine translation system, but it accomplishes this without updating MT engines, which should make it easier to implement for additional languages.

Pairaphrase's Interface Resembles a Simple Version of a CAT Tool
Source: Pairaphrase
 

Company founder Rick Woyde emphasizes that his target customers would not normally invest in language technology, but see the value in assistive communication technologies when working in other languages. Like the Japanese company Yaraku Zen, which targets a similar market, Pairaphrase is betting on the potential of professionals in other fields who speak two or more languages and would benefit from MT as an aid to speed up their work across language boundaries.

A Thousand AI Flowers Will Bloom
Although the companies discussed above have very different emphases, what unites them is a willingness to apply AI to new fields. Trained artificial intelligence engineers might be in short supply, but the basics building blocks of AI are increasingly accessible and encourage innovators to find instances where it can reduce manual work or open new markets. As the technology advances, expect to see more and more developers and LSPs looking for ways to eliminate redundant and mundane tasks, many of them in areas we have yet to anticipate.