On May 12, 2020, Lilt announced that it had raised USD 25m in its latest funding round. This brings the total funds raised to USD 37.5m — making Lilt one of the most well-funded language industry startups of the past 10 years after Smartling and Unbabel.
“We closed the round several months ago, but we announced it on Tuesday, May 12,” CEO Spence Green told Slator. He said that while Covid-19 had no impact on the financing, they did postpone the announcement as a result of the pandemic. Securing the funding before Covid-19 disrupted the global economy was fortunate, as the funding environment has since fundamentally changed.
The Series B funding round was led by Intel Capital and joined by existing investors Sequoia Capital, Redpoint Ventures, In-Q-Tel, Zetta Venture Partners, and XSeed Capital, according to a press statement; which also stated that Intel Capital’s VP and Senior Managing Director Mark Rostick joined Lilt’s board of directors.
At SlatorCon San Francisco 2019, Redpoint Ventures Partner and Managing Director, Tomasz Tunguz, discussed their rationale for investing in Lilt based on their thesis of the AI Agency (accounting firms, law firms, debt collection, language service providers, etc.) replacing the traditional agency business model.
From SaaS to Enterprise
Co-founders John DeNero and Spence Green started Lilt in 2015 as a translation productivity tool targeted at individual translators; that is, using adaptive machine translation, which allows the text to change dynamically, suggesting next words or phrases as the translator types.
Lilt then expanded their service offering to cater to language service providers (LSPs) and, in 2018, pivoted away from a subscription-based model to take on enterprise clients. As Lilt became a full-on tech-enabled LSP, the San Francisco-based company secured USD 9.5m in funding.
The fresh funds did not all go into hiring dozens of Computer Scientists and Marketers though. The company also understood that to expand within the enterprise, they needed people experienced in the complex sales, onboarding, and account management required for large enterprise accounts.
So, in January 2019, Lilt announced the addition of former Lionbridge Chief Sales Officer, Paula Shannon, as an advisor. Within months, Shannon brought on fellow Lionbridge veteran Roberto Sastre to be Head of Revenue EMEA.
The 3 Most Exciting Trends in MT
As the rapid pace of research in the broader machine translation (MT) space continues unabated, with dozens of new papers published on a weekly basis, it is sometimes hard to parse through what is of purely academic interest and what has real-life practical applications. So we asked which areas Computer-Scientist-turned-CEO Green regarded as the most exciting in current MT research. He singled out three.
1. Exceptional improvement in personalized MT models. “MT models that are adapted to the particular way in which a translator or organization uses language continue to improve dramatically. Now, by changing only a small fraction of the model’s parameters, we can be very effective in personalizing a very large and high-performance, general neural machine translation model. What this means is that a service like Lilt can use larger models with better out-of-the-box performance — and still train personalized models for every customer, every translator, and even every individual document.”
“We’ve also spent a lot of time working to reduce latency to make sure that translation suggestions keep up with typing speed” — Spence Green, CEO, Lilt
2. Ability to add more contextual information in translation suggestions. “We’re discovering how to include even more contextual information when making translation suggestions. Document context, project-specific terminology, and translation memory matches will all be part of the input to machine translation in a way that’s more flexible and tightly integrated than past solutions. The result is that terminology will be inserted in the right place with the right inflection, and translation suggestions will be coherent across the different segments of a document.”
3. More efficient automatic tag placement. “Our recently published work (accepted at ACL 2020) on end-to-end neural word alignment shows that MT models can be much more effective at automatic tag placement than was previously known — which will let linguists focus more of their time on translation and less on the drudgery of placing tags. We expect further advancements in this direction in the coming years.”
Reducing Latency
On how their adaptive neural machine translation (NMT) interface has progressed since they discontinued the Lilt PRO software license and closed the platform, Green said, “Our most recent updates have improved overall translation quality, greatly improved robustness to things like capitalization and non-translatable elements, and increased the speed that a model adapts to a linguist’s style and language use.”
Moreover, according to Green, “Because our platform is cloud-based, we’ve also spent a lot of time working to reduce latency to make sure that translation suggestions keep up with typing speed.”
The Lilt CEO also shared the areas where they are likely to deploy funds in the near-term. “We are focused on investing these funds in four areas: customer enablement, research and product, our service model, and our people.”
Elaborating on plans for Lilt’s service model, Green said, “To increase supply-chain flexibility we’re investing in new offices for broader timezone coverage.”
An expansion in physical office presence makes sense for Lilt. While much of the language industry’s operational model is well suited for remote work, selling to and supporting major enterprise accounts still requires a certain in-country presence.