We ran a two-week A/B test. Same short drama series. Pure human: 8 episodes a day. Pure MT: 24 episodes a day. MTPE with a trained post-editor: 16 episodes a day, and the quality was 88 out of 100.
The pure human translation scored 92. The pure MT scored 54. The MTPE workflow hit 88 while doubling the pure-human throughput.
I want to be clear about what that 88 means and what it doesn't mean, because numbers without definitions are just decoration. The quality score in this test was a composite: grammatical accuracy, idiomatic naturalness, character voice consistency, cultural adaptation, and subtitle timing compliance. These were weighted. Grammatical accuracy got 20% of the score. Character voice consistency got 25%, because in short drama translation, characters sounding different from each other is more important than perfect grammar. Cultural adaptation got another 25%, because a line that's grammatically perfect and culturally tone-deaf is still a failed line.
The pure MT failed hardest on character voice and cultural adaptation. It was grammatically fine. Modern neural MT is grammatically fine most of the time. But every character sounded the same, and the cultural nuances that make short drama dialogue work — the sarcasm, the power dynamics, the unspoken tension — were flattened into informational text.
The MTPE workflow kept the speed advantage of MT for the straightforward content and deployed human editorial judgment on the content that actually needed it. That's the insight. Not 'use AI for everything and have a human check it.' That's not MTPE. That's AI with a rubber stamp. The actual MTPE value proposition is knowing which content needs human intervention and which content doesn't, and applying the right level of editorial effort to each line.
The dead loop: why 'just hire more translators' stopped working for short drama
Short drama platforms are operating on content calendars that would have been science fiction five years ago. A typical platform might be releasing 20-30 new episodes per day across multiple series. Each episode is 1-3 minutes. The dialogue density varies wildly — some episodes are mostly visual, some are dialogue-heavy confrontation scenes — but you can't predict the density distribution when you're scheduling the translation pipeline.
The math breaks down fast. A competent subtitle translator working on short drama content can handle roughly 20-25 minutes of runtime per day at acceptable quality. That's maybe 8-12 episodes depending on length. If the platform needs 30 episodes translated per day, pure human translation means either hiring three translators per language pair — which is expensive and difficult for many language pairs — or accepting that episodes will go untranslated and viewers will churn.
Pure MT solves the speed problem. Modern neural MT engines can process an episode's dialogue in seconds. You can run 30 episodes through MT in under ten minutes. But the output quality on short drama content — which is dialogue-heavy, emotionally loaded, and full of culturally specific references — is not broadcast-ready. I don't care how good the MT engine is. That gap between 'informational accuracy' and 'dramatic effectiveness' is where the human editorial judgment belongs.
So you've got human translation: slow but good. MT: fast but flat. The gap between them is where MTPE lives, and designing an MTPE workflow that actually delivers something better than either approach alone is the problem I want to walk through.
What MT actually gets wrong on short drama content, specifically
I've spent enough time reviewing MT output on short drama dialogue to have a pretty clear taxonomy of failure modes. These aren't generic MT problems. They're specific to this content type.
Emotional register flattening. MT translates words. It doesn't translate intensity. A Chinese line like '你给我滚' — which, depending on context and delivery, can range from 'get lost' to 'get the hell out of my sight right now' — will almost always come out of MT as something like 'you go away' or 'leave.' The words are technically there. The emotional payload is gone. And in short drama, the emotional payload is the content. Nobody watches short drama for the plot architecture. They watch for the emotional beats.
Character voice erasure. A short drama might have a cold CEO who speaks in clipped, formal sentences, a sassy best friend who uses slang and sentence fragments, and a villain who speaks in elaborate metaphors. The MT engine sees three instances of dialogue and translates all three in the same neutral register. The cold CEO and the sassy best friend end up with identical sentence structures and identical vocabulary. The viewer can't tell who's speaking from the subtitles alone. This isn't a minor quality issue. In a medium where dialogue is the primary vehicle for character, character voice erasure is a fundamental content failure.
Idiom and slang collapse. MT handles idioms one of two ways: literal translation or generic replacement. Chinese '吃醋' (literally 'eat vinegar,' meaning jealousy) becomes either 'eat vinegar' (literal nonsense) or 'jealous' (accurate but flat). A human translator might render it as 'green with envy' or 'she's got a jealous streak' or 'someone's feeling territorial' depending on context, register, and character. MT can't make that call because it doesn't understand context, register, or character.
Cultural allusion blindness. Short dramas are packed with culturally specific references — historical allusions, pop culture callbacks, social media in-jokes, regional idioms. MT either translates these literally (nonsense to the target audience) or replaces them with the closest statistical match in its training data (which is usually wrong because the closest statistical match was learned from a completely different context). A human translator knows that a reference to a specific Chinese historical figure needs to be either adapted to a culturally legible equivalent or reworked to convey the narrative function without the specific reference.
Subtitle timing blindness. MT produces text. It doesn't know that the text is going to be displayed on screen for a specific duration at a specific position relative to shot changes. A beautiful MT translation that runs 90 characters for a line that has 2.1 seconds of screen time is unusable. The post-editor has to compress it to ~35 characters while preserving meaning, which is sometimes more work than translating from scratch would have been.
These five failure modes define where the post-editor's time should go. The MT is fine for the straightforward stuff — 'meet me at the cafe at 3 PM,' 'the contract is on your desk,' 'I'll call you later.' The post-editor's attention should be directed at the lines where emotional register, character voice, idiom, or cultural reference is in play.
Light edit vs. deep edit: the tiered approach that actually works
The most common MTPE mistake I see is treating every line with the same editorial intensity. Some lines need a complete rewrite. Some lines are perfectly fine as-is. Treating all lines the same way wastes post-editor time on the lines that don't need it and gives insufficient attention to the lines that do.
The tiered approach that I've seen work consistently in short drama MTPE workflows:
Tier 1: No-touch lines (roughly 40-50% of dialogue). These are informational lines where the MT output is accurate, idiomatic enough for subtitles, and doesn't carry character-critical or emotion-critical content. 'The meeting is at 3 PM.' 'I sent the file.' 'She's waiting in the lobby.' The post-editor skims these — confirmation glance, no edits, move on. Time per line: 1-2 seconds.
Tier 2: Light-edit lines (roughly 30-40% of dialogue). These lines carry emotional or character content but the MT output is directionally correct — it just needs tuning. The MT rendered a sarcastic line as a neutral statement. The vocabulary is correct but the register is wrong for the character. The line is too long for the subtitle timing window and needs condensation. The post-editor adjusts register, compresses for timing, sharpens emotional tone. Keeps the MT output as the base but sculpts it. Time per line: 5-15 seconds.
Tier 3: Deep-edit lines (roughly 10-20% of dialogue). These are the lines where MT failed completely — emotional climaxes, culturally specific references, idiom-heavy dialogue, character-defining moments. The MT output is either nonsensical or so flat that using it as a base would be counterproductive. The post-editor essentially retranslates these lines from scratch, using the MT output only as a reference for vocabulary and sentence structure that can be discarded. These lines might represent only 15% of the total dialogue but they represent probably 50% of the viewer's emotional experience of the episode, because these are the big moments. Time per line: 20-60 seconds.
This tiered system is what makes the 50% speed improvement possible. The post-editor is spending almost no time on lines that don't need editing, moderate time on lines that need tuning, and significant time on lines that need full intervention. The total editing time per episode is roughly 50-60% of what a full human translation would take, because the post-editor isn't translating from scratch — they're editing from a base that's mostly correct for most of the content.
The critical design decision is the threshold between tier 1 and tier 2. If you classify too aggressively into tier 1, content that needed light editing goes unedited. If you classify too conservatively, you're applying light edits to content that didn't need them, which erodes the speed advantage. Training post-editors on the classification criteria — essentially teaching them to recognize emotional register, character voice, and cultural reference density at a glance — is the single most important investment in making this workflow effective.
What training a post-editor actually involves
I want to talk about this because there's a persistent assumption in the industry that MTPE is 'translation but easier' and requires less skilled people. That assumption is wrong in a way that produces terrible output.
A good MTPE post-editor for short drama content needs to be a better translator than a pure human translator doing the same content, for a very specific reason. A pure human translator starts from the source text and produces the target text. The cognitive workflow is: source → comprehension → target. An MTPE post-editor starts from the MT output and has to evaluate whether it's accurate, then decide whether to keep, modify, or discard it, then produce the target text. The cognitive workflow is: source → MT output → evaluation → decision → target. That's two extra cognitive steps per line — evaluation and decision — and the post-editor has to execute them fast enough to maintain the throughput advantage.
This means the post-editor needs: native-level target language proficiency, obviously. Strong source-language comprehension, also obvious. But they also need enough experience with short drama content specifically to recognize at a glance which lines are tier 1 vs tier 2 vs tier 3. They need to understand subtitle timing constraints well enough to make compression decisions on the fly. They need to be familiar with the specific MT engine's failure modes — every MT engine has patterns of what it gets wrong, and an experienced post-editor learns those patterns and starts looking for those specific error types automatically.
And critically, they need the editorial confidence to say 'this MT output is wrong and I'm rewriting it from scratch,' because there's a psychological pressure in post-editing to preserve the MT output. It feels wasteful to discard perfectly formed sentences. The metric of MT edit distance — how much the post-editor changed from the MT output — is sometimes used as a quality metric for MTPE, and it's actively harmful in short drama content because the lines that need the most editing are the high-impact lines where preserving the MT output would do the most damage.
Training a post-editor for this workflow takes about two weeks of supervised practice with feedback, assuming they're already an experienced subtitle translator in the language pair. The training content is real episodes with annotated MT output — here's the MT, here's how a senior post-editor classified each line, here's what the senior editor changed and why. The post-editor practices on subsequent episodes and gets feedback on their classification decisions and edit quality. After about 20-30 episodes of practice, most experienced translators can hit the tier classification accuracy and edit quality targets.
The quality control layer that prevents MTPE from becoming 'MT with a quick glance'
MTPE workflows tend to collapse toward one extreme or the other over time. Either the post-editor starts over-editing everything because it feels safer (which kills the speed advantage), or they start under-editing because the deadline pressure is intense (which kills the quality advantage). The only thing that prevents either collapse is a structured quality control layer.
What I've seen work:
Spot-check review on a per-episode basis. A senior reviewer spot-checks roughly 10-15% of each post-edited episode, weighted toward tier 2 and tier 3 lines. The reviewer is checking for: classification accuracy (were tier 3 lines correctly identified?), edit quality (do tier 2 edits actually improve the MT output?), and error patterns (is the post-editor consistently missing a specific type of MT failure?). The review takes about 15-20 minutes per episode and catches the majority of systematic quality issues before they propagate across multiple episodes.
Post-editor calibration sessions. Every 20-30 episodes, the post-editing team does a calibration session: same episode, independently post-edited by multiple editors, reviewed together to identify classification and editing differences. This prevents editorial drift, where each post-editor gradually develops their own internal standard for what constitutes tier 2 vs tier 3 and the output becomes inconsistent across editors. The calibration session takes about two hours and it's the most effective quality investment I know of in an MTPE workflow.
MT engine-specific error logs. The post-editing team maintains a running log of MT failure patterns specific to the engine and language pair they're working with. 'Engine consistently renders sarcasm as neutral statement.' 'Engine consistently mistranslates Character X's honorific speech patterns.' This log becomes a reference for new post-editors joining the team and a training tool for improving classification speed. It also feeds back to the MT engine training team, if the platform has one.
These three QC mechanisms are not expensive relative to the cost of the alternative, which is either: pure human translation at half the throughput, or pure MT at less than half the quality. The investment is in process design and team training, not additional headcount.
When MTPE doesn't work for short drama (and what to do instead)
I should say this because I don't want anyone walking away thinking MTPE is the answer to every short drama translation problem. It's not.
MTPE falls apart when the content is too culturally dense for the MT engine to produce a usable base. If the show is set in a specific historical period with period-specific language conventions, the MT will fail so consistently on register and terminology that the post-editor is effectively translating from scratch on 60%+ of lines. The speed advantage disappears. Use pure human translation for historically dense content.
MTPE falls apart when the language pair is low-resource and the MT engine's training data is insufficient. If the MT is producing grammatically broken output on a regular basis, the post-editor is spending more time fixing grammar than they would spend translating from scratch. The speed advantage inverts — MTPE becomes slower than pure human translation. This is especially common for language pairs where neural MT training data is sparse.
MTPE falls apart when the post-editor isn't given enough time to do tier 3 lines properly. If the per-episode editing budget is so tight that the post-editor can't spend 20-60 seconds on the 10-20% of lines that need deep editing, the emotional climaxes of the episode are going to fall flat. MT output on emotional climaxes is almost always bad because MT doesn't understand dramatic stakes. If your MTPE budget doesn't allow for deep editing on the big moments, you're better off doing pure human translation on those moments and MT on the rest.
The decision to use MTPE should be content-driven, not budget-driven. Look at the actual content. Look at the actual MT output for that content and language pair. Run a test episode through the tiered MTPE workflow and compare the output quality and throughput to both pure human and pure MT baselines. If the MTPE output isn't clearly better than both alternatives in the speed/quality tradeoff space, don't use it. Use whatever actually works for the specific content and language pair you're dealing with.
Artlangs Translation builds MTPE workflows for short drama content: trained post-editors who understand dramatic dialogue, tier-based editorial systems (no-touch, light edit, deep edit), structured spot-check review, calibration sessions to prevent editorial drift, and content-driven decision-making about when MTPE is and isn't the right approach. 230+ language pairs. If your short drama translation is stuck in the 'fast but bad' vs 'good but slow' dead loop, the MTPE tiered workflow is how you break out of it — but only if the post-editors know what they're editing and why.
