Should You Trust NMT or CAT? A Post-Editor's Field Guide

Imagine you're a professional translator staring down a 5,000-word legal contract at 4 PM, deadline looming. Your CAT tool is open, translation memory loaded, but the client has also provided raw machine translation output and asks, 'Can you just post-edit this?' Your gut says no—it'll take longer to fix the MT than to translate from scratch. But is that true? And what about the new neural engines claiming near-human quality? The translation industry is in a weird spot: neural machine translation (NMT) has become shockingly good, yet the tools we use to produce professional work—CAT tools—haven't changed all that much. So, the real question you need answered is this: Should you embrace NMT as a post-editing crutch, or stick to your CAT-tool guns? This isn't a philosophical debate; it's a practical, hourly-rate decision. Let's dig in.

The Core Difference: MT Replaces You, CAT Assists You

Let's get the basics straight because the confusion is widespread. Machine translation (MT) is fully automated—it converts text between languages without a human in the loop, using rule-based, statistical, or neural approaches (Machine translation, Wikipedia). On the other hand, computer-assisted translation (CAT) tools are designed to aid human translators, not replace them; they leverage translation memory (TM), a database of previously translated segments, to boost consistency and cut down on repeated work (Machine translation, Wikipedia). In short: MT is the autopilot, CAT is the power steering. When you use a CAT tool, you're in control; when you let MT run, you're just the cleanup crew. That distinction matters because it changes your workflow, your liability, and your value.

Why NMT Is Not Your Replacement (Yet)

Neural machine translation, the tech behind Google Translate and DeepL, is a marvel. It analyzes whole sentences rather than word-by-word, and it's trained end-to-end on massive parallel corpora (Machine translation, Wikipedia). When Google launched its GNMT in September 2016, it reduced translation errors by 55–85% on several major language pairs compared to the previous phrase-based system (GNMT at production scale, Google Research). DeepL, launched in 2017, went further, claiming that professional translators chose its output roughly three times more often than that of Google, Microsoft, or Facebook in blind tests (DeepL Translator, Wikipedia). Those numbers are impressive. But here's the catch: NMT still stumbles on idiomatic expressions, cultural nuance, and domain-specific terminology, which is why post-editing remains essential (Machine translation, Wikipedia). The ISO even has a standard—ISO 18587:2017—that specifies requirements for full human post-editing of MT output and the competences post-editors need (ISO 18587:2017, ISO). That standard exists because MT output, even the best, isn't publishable without a human. So, if you're a professional translator, NMT isn't going to eat your lunch, but it might eat your post-editing rate if you don't price it right.

CAT Tools: Your Consistency Workhorse

CAT tools aren't just a safety net; they're a productivity multiplier. A translation memory stores your previous work, so when a segment repeats or matches, you get it for free. That's why tools like SDL Trados, memoQ, and Memsource are industry staples—they're built around TM (Machine translation, Wikipedia). But here's the thing: CAT tools have been around for decades, and their core functionality hasn't changed much. Google Translator Toolkit, which ran from 2009 to 2019, was a CAT tool that split documents into segments, pre-translated them from TM, and fell back on MT when no match was found (Google Translator Toolkit, Wikipedia). That's the same pattern you see in modern CAT tools with MT integration. The difference now is that the MT fallback is neural, not statistical, and it's a lot more useful. But the tool itself is still a database with a bilingual editor. So, if you're hoping a CAT tool will magically make you faster, it won't unless you're using it to leverage your own memory and glossaries. The real value of CAT is consistency, not speed.

Choosing Your Stack: The NMT-CAT Hybrid

So, what's the right approach? You need both, but you need to use them in the right order. Here's my blunt recommendation: use a CAT tool as your primary workspace, and integrate NMT as a pre-translation layer, but only for segments that don't match your TM. That way, you're not reinventing the wheel for new content, but you're also not letting MT run wild on segments you've already translated. The table below breaks down the options:

The hybrid approach isn't new—Google Translator Toolkit did it back in 2009—but the neural engine changes the game. For example, if you're translating a user manual with a lot of repeated boilerplate, your TM will cover maybe 60% of the segments. The remaining 40% can be pre-translated with NMT, and you post-edit. In my experience, post-editing good NMT output takes about half the time of translating from scratch, but that's only true for clean, well-written source text. If the source is messy, NMT will amplify the mess.

Metrics: Why BLEU Is Lying to You

You might be tempted to evaluate an MT engine using BLEU scores. Don't. BLEU, introduced by IBM in 2002, compares n-gram matches between machine output and human references, scoring on a 0–1 scale (IBM Research). It's simple and cheap, but it correlates poorly with human judgment (IBM Research). In fact, IBM Research itself now says BLEU is increasingly seen as unreliable, and newer metrics like COMET and BLEURT are becoming standard (IBM Research). COMET, developed by Unbabel, is a neural framework that won the WMT 2019 and 2020 metrics shared tasks (COMET, ACL Anthology). BLEURT, from Google, is BERT-based and correlates better with human judgment on NLG tasks (BLEURT, Google Research). So, when you're choosing between engines or tuning your post-editing workflow, don't chase BLEU. Look for human evaluation, or use COMET-style metrics if you must automate. And if you're pricing post-editing, use HTER—Human-targeted Translation Edit Rate—which measures the edits a human makes to MT output, a better proxy for post-editing effort (Translation Edit Rate, ACL Anthology).

The Bottom Line: Your Workflow, Your Rules

Here's the actionable takeaway. Stop thinking of MT and CAT as rivals. They're different tools for different jobs. Use a CAT tool for consistency and TM leverage. Use NMT for speed on new content, but only as a pre-translation that you post-edit. And when you post-edit, charge for it—post-editing is skilled work, not proofreading. The ISO 18587 standard exists because post-editing requires specific competences (ISO 18587:2017, ISO). So, if a client asks you to 'just run it through Google and fix the errors,' push back. Explain that you'll use your CAT tool with TM, then NMT for new segments, then post-edit to publishable quality. That's a workflow that respects your time, your client's budget, and the language itself. If you don't, you'll be doing the work of a machine and getting paid like one.

Quick tip: Always match the engine to the content. For legal or medical text, even the best NMT will need heavy post-editing; for marketing copy, you might as well translate from scratch. Don't let a cheap MT output dictate your process.

In the end, the question isn't whether to trust NMT or CAT. It's whether you're willing to adapt your workflow to get the best of both. The industry is changing—the EU's Directorate-General for Translation produces 2.6 million pages a year, and they use eTranslation, their free neural MT service, as a productivity aid (European Commission translation department; eTranslation, European Commission). If the biggest translation service in the world uses MT to boost human output, you should too. But you have to keep the human in control. That's your value. Don't give it away.

Sources

• Machine translation - https://en.wikipedia.org/wiki/Machine_translation
• ISO 18587:2017 - https://www.iso.org/standard/62970.html
• IBM Research - https://research.ibm.com/blog/bleu-nlp-benchmark-anniversary
• COMET - https://aclanthology.org/2020.emnlp-main.213/
• Google Research - https://research.google/blog/a-neural-network-for-machine-translation-at-production-scale/
• European Commission translation department - https://commission.europa.eu/about-european-commission/departments-and-executive-agencies/translation_en