Human‑Interactive Machine Translation: Research, Techniques, and Productization

The talk begins with an overview of machine translation (MT) development, noting that many companies adopt MT to showcase AI capabilities despite low translation demand and persistent quality issues such as ambiguity, unknown terms, and non‑literal expressions.

It describes the dominant encoder‑decoder framework, the evolution from RNN‑based models to attention mechanisms and the Transformer architecture, highlighting how self‑attention enables richer contextual encoding at the cost of higher computational resources.

Evaluation metrics such as BLEU and perplexity (PPL) are explained, and the need for large‑scale data and GPU clusters for training state‑of‑the‑art models is emphasized.

The article then introduces human‑interactive MT, defining three core tasks: user‑guided translation interventions, real‑time learning from corrections, and provision of auxiliary translation information, illustrating how human feedback can improve model outputs.

Practical applications at Tencent are outlined, including simultaneous interpretation, image‑based translation, and assisted translation tools, with discussion of internal versus external deployment scenarios and the importance of aligning technical solutions with product requirements.

Finally, the author reflects on AI productization, stressing the need for multidisciplinary teams (researchers, engineers, product managers), the challenges of data acquisition, open‑source integration, hardware constraints, and the strategic decision between building a "AI product" versus embedding AI into existing products.