Spence Green — Enterprise-scale Machine Translation

Early Days

The early days of machine translation were marked by naive assumptions and rudimentary systems. Spence Green explains that initial efforts in the 1950s treated languages as simple ciphers, leading to ineffective translation systems 1. These early systems relied on a linear editing workflow, where human intervention was necessary to correct machine outputs. Green notes that this approach was inefficient, prompting the development of predictive typing interfaces and online learning to improve translation accuracy 2.

The naive idea was, well, let's just take the machine output and pay somebody to fix it.

--- Spence Green

This evolution laid the groundwork for more sophisticated machine translation systems.

   

Neural MT

The transition to neural machine translation (NMT) marked a significant leap in translation quality. Spence Green highlights how NMT transformed machine translation from "funny" to "meaningfully good," revolutionizing the field 3. The adoption of frameworks like TensorFlow allowed for more efficient system architectures, reducing code complexity and enhancing performance 4.

MT just went from being kind of funny to being meaningfully good.

--- Spence Green

This shift required rapid adaptation and skill acquisition, as companies like Lilt built new systems from scratch to keep pace with advancements.

   

Technical Hurdles

Developing machine translation systems involves overcoming numerous technical challenges. Spence Green shares that transitioning from research prototypes to production-ready systems was unexpectedly complex, taking nine months instead of the anticipated six weeks 5. Challenges included managing latency constraints and ensuring efficient CPU inference for personalized models.

The amount of detailed large scale engineering work that has to go into that was surprising to us.

--- Spence Green

These hurdles highlight the intricate engineering required to deliver high-quality translations at scale 1.