When you run a reverse-Taobao storefront that sells to buyers in Korea, Japan, Southeast Asia, and the English-speaking world at the same time, product copy becomes a real engineering problem. Machine translation gets you 80% of the way, then quietly breaks the last 20% where it matters most.
Taocarts, a cross-border system built for agents and resellers sourcing from Taobao and 1688, supports multiple languages out of the box. But "supports multiple languages" is not the same as "every field is safe to auto-translate." We learned that the hard way after a wave of avoidable returns.
The trap of translating everything
Early on we pushed every field—title, specs, size chart, return policy—through an MT engine and called it localized. Returns spiked. The reason was boring: size charts translated literally, and material descriptions lost the nuance that actually decides a purchase. MT is fine for descriptive filler; it is dangerous for anything that changes fit or liability.
Drawing the boundary
We split listing fields into three buckets:
- MT-safe — marketing blurbs, general descriptions, category names.
- MT-with-glossary — brand names, material terms, measurement units. Translated but locked to a glossary so "cm" never drifts inconsistently.
- Human-only — size charts, care instructions, compliance and return statements.
The third bucket is small but high-stakes. A human reviewer signs off before publish. This is not about distrusting the model; it is about containing the blast radius.
How it works in the data model
Each listing carries a source_lang and a set of target_lang copies. The MT pipeline fills buckets 1 and 2 automatically and flags bucket 3 for review. A simple status field—draft / mt_ready / human_review / published—drives the whole workflow. Reviewers see source and MT side by side and only touch what needs touching.
Why it pays off
The win is not "better translation." It is fewer disputed orders and a review queue that stays tiny because most content never reaches a human. After drawing this line, listing-related returns roughly halved for us.
Localization is not a translation problem you throw a model at. It is a routing problem: decide what a machine may touch and what a person must own, then enforce it in the schema. Do that, and rolling out a new language market stops being the scary part of the quarter.
