My Role
Webinterpret provides comprehensive cross-border e-commerce solutions for eBay and Amazon sellers, offering a SaaS platform that enables them to launch their stores in over 10 markets.
As a Product Manager at Webinterpret I took the lead on development of the key component of Webinterpret’s end-to-end cross-border solution for international sales on Amazon and Ebay.
In my role, I defined clear requirements for the new translation system and the initial proof of concept (POC), while also managing weekly reports to the CEO and the VP of engineering. I discussed implementation details, potential risks, and scalability with the engineering team, and established a testing framework with our internal team of linguists.
Challenge
My objective was to optimize Webinterpret’s translation services by reducing costs by 70% within a year to enable the company to scale and onboard more users without increasing translation budget.
The challenge of this task laid in the system’s complexity, where each listing element—title, category, description, bullet points, and attributes—was processed either through different machine translation modules or by human translators.
Solution
To thoroughly understand the system, I collaborated with a senior engineer to create detailed documentation. This effort allowed me to identify that approximately 10% of the content, which accounted for about 98% of translation costs, was handled by human translators. These costs varied significantly across categories (like clothes, beauty etc).
After analyzing the data, I decided to enrich our dictionaries, aiming to reduce costs by 50%-80% in these areas. The initial results were promising; however, translation costs began to increase again due to constantly changing inventory and new vocabulary. Further analysis led me to a conclusion that optimizing existing algorithms was insufficient due to the long tail of vocabulary distribution.
I presented this problem to our data scientists, and we collectively decided to explore alternative solutions that could adapt to the evolving vocabulary. Our lead data scientist suggested a statistical machine translation (SMT) approach based on language models. After reviewing several publications and discussing the concept with our data science and engineering teams, I was convinced of its potential. I prepared a business case, demonstrating that continued investment in the existing system would not sustainably reduce costs. With approval from the head of engineering and the CEO, we began developing statistical models trained on extensive translation examples.
Once the system’s quality and performance reached satisfactory levels, I decided to launch it for a select group of users and listings. After successful initial tests, I received approval from the VP of engineering and the CEO to transition all users to the new system. I continued to monitor quality and costs closely and engaged with the customer support team to track any issues related to translation quality or performance.
Results
As a result, my team and I not only met but exceeded our annual goals. We managed to translate over 10 million more listings without increasing the translation budget. The system was patented and continues to be utilized by Webinterpret to translate listings for more than 33,000 sellers globally.