How is Translation Edit Rate calculated?

TER counts the single-word edits needed to change the hypothesis into the reference — insertions (I), deletions (D), substitutions (S), and block shifts (Sh) — then divides by the number of reference words: TER = (I + D + S + Sh) / N_ref. Every edit, including a shift of a block of any length, costs exactly 1. It is computed with a greedy shift search followed by a word-level Levenshtein distance, per Snover et al. (2006).

What is a good TER score?

Lower is better, and 0% means the output already equals a reference. There is no official cut-off, but as an informal guide: under 10% is excellent, 10–30% good, 30–50% fair, and over 50% usually means heavy post-editing. TER can exceed 100% when the hypothesis is much longer than the reference, because edits are divided by the reference length, not the hypothesis length.

Is a lower or higher TER score better?

Lower. TER measures effort — the fraction of the reference that would have to be edited — so a smaller number means the machine output is closer to the human reference. This is the opposite of BLEU, chrF, and METEOR, where higher is better. A TER of 0 is a perfect match; a TER of 1.0 (100%) means roughly one edit per reference word.

What is the difference between TER and WER?

Word Error Rate (WER) is a plain word-level edit distance divided by the reference length — insertions, deletions, and substitutions only. TER adds block shifts: reordering a phrase costs one edit instead of being paid for as several substitutions or a delete-plus-insert. So for a correct translation that is simply in a different word order, TER is much lower than WER. Turn off 'Count shifts' here to see the WER-style number.

Does TER count word reordering as an error?

Yes, but cheaply. Moving a contiguous block of words that matches the reference to its correct position counts as a single shift edit, no matter how long the block is. This is the key idea from Snover et al. (2006): a reordering that a plain edit distance would punish as many substitutions is charged only once, which better reflects the real post-editing effort.

What is the difference between TER and BLEU?

BLEU rewards n-gram overlap and returns a higher-is-better score; TER counts edits and is lower-is-better. BLEU is usually reported at the corpus level, while TER is naturally a per-sentence effort estimate. They often disagree on individual sentences, so many machine-translation papers report both. This tool scores a single sentence pair; use the BLEU calculator for the overlap view.

Does this match sacrebleu?

It follows the same tercom algorithm sacrebleu implements. The default settings (lowercase on, normalize punctuation on) reproduce sacrebleu with normalization enabled; turn normalization off to match sacrebleu's plain --metrics ter default. Tokenisation and case affect the score, so match those settings when comparing numbers. The three worked examples below reconcile with sacrebleu exactly.

Can I score against more than one reference?

Yes. Put one reference per line in the reference box. The tool scores the hypothesis against each reference independently and reports the minimum (best) TER, normalised by that reference's length — the same multi-reference behaviour as tercom. The results panel shows which reference won.

Does my text get uploaded anywhere?

No. The whole calculation runs in your browser with JavaScript — nothing is sent to a server, logged, or stored. You can paste confidential or unpublished translations safely and even use the tool offline once the page has loaded.

AI · Machine Translation

TER Score Calculator (Translation Edit Rate)

Paste a hypothesis and one or more references to get the Translation Edit Rate — the fraction of reference words you would have to edit — with a full insertion, deletion, substitution, and shift breakdown plus a word-level alignment. Lower is better. Matches sacrebleu, no signup, runs in your browser.

By Induwara Ashinsana— Executive Director, Ryzera TechnologiesUpdated Jul 9, 2026

TER Score

Hypothesis (machine output)

One sentence. Stays on your device.0 words

Reference(s) — one per line

At least one. Stays on your device.0 words

Options

Examples

TER

—

Rating

—

Total edits

—

Reference length

—

Runs entirely in your browser — your text is never uploaded, logged, or stored. Method: greedy block-shift search plus word-level edit distance, each edit costing one, per Snover et al. (2006); reconciled to sacrebleu --metrics ter. Up to 20,000 characters per box.

How it works

TER (Translation Edit Rate) estimates how much work it would take to turn a machine hypothesis into a correct human reference. Unlike overlap metrics such as BLEU or chrF, it is an error rate, so a lower number is better and 0% is a perfect match. It was defined by Snover et al. (2006) and is the metric the tercom and sacrebleu tools compute. The formula is:

TER = (I + D + S + Sh) / N_ref

where I is insertions, D deletions, S substitutions, Sh shifts, and N_ref the number of reference words. Every edit costs exactly one — including a shift of a contiguous block of any length. It is computed in four steps:

Tokenise and normalise.Both strings are split on whitespace into word tokens. With the defaults, text is lowercased and punctuation is separated into its own tokens, following tercom's normalisation (the exact sacrebleu TercomTokenizer regexes).
Greedy shift search. The tool repeatedly looks for a single shift — moving a contiguous hypothesis block that exactly matches a reference word sequence to a new position — that most reduces the plain word-level edit distance. The best-improving shift is applied, counted as one edit, and the search repeats until no shift reduces the distance. With Count shifts off, this step is skipped, giving a plain edit rate (the same idea as Word Error Rate).
Residual edit distance. After shifting, the remaining insertions, deletions, and substitutions are counted with a standard Levenshtein alignment between the shifted hypothesis and the reference.
Score.Add the shift count to the residual edits and divide by the reference length. With several references, each is scored independently and the minimum (best) TER is reported, normalised by that reference's length — matching tercom's multi-reference behaviour.

Because finding the truly minimal edit sequence with shifts is computationally hard, TER uses this greedy heuristic, exactly as the reference implementations do. As a safeguard, the residual edit count is derived two independent ways — the dynamic-programming distance and a tally of the aligned operations — and the score is flagged “cross-checked” only when they agree. The numbers reconcile with sacrebleu --metrics ter when tokenisation and case settings match.

Worked examples

Substitution + deletion → 33.33%

Hypothesis: the cat sat on mat
Reference: the cat is on the mat

Align: the=the, cat=cat, is→sat (1 substitution)
on=on, then the reference 'the' is missing (1 insertion)
mat=mat. No block shift reduces the edit distance
Edits = 1 sub + 1 ins = 2 over N_ref = 6
TER = 2 / 6 = 0.3333 → 33.33%

Pure reordering, one shift → 16.67%

Hypothesis: to the school the boy went
Reference: the boy went to the school

The block 'to the school' matches the reference tail
Shift it to the end → 'the boy went to the school'
That equals the reference: 1 shift, 0 residual edits
TER = (1 + 0) / 6 = 0.1667 → 16.67%
Turn shifts off: no monotonic alignment matches both swapped blocks, so all 6 words need editing → 6/6 = 100%

Insertion only → 25%

Hypothesis: she will arrive
Reference: she will arrive tomorrow

she=she, will=will, arrive=arrive (3 matches)
The reference word 'tomorrow' is missing from the hypothesis
That is 1 insertion; no shift helps
N_ref = 4 reference words
TER = 1 / 4 = 0.25 → 25.00%

Frequently asked questions

Sources & references

The formula and the worked examples on this page were last reconciled by hand and against sacrebleu --metrics ter on 2026-07-09. The calculation module ships with a built-in assertion that re-runs every worked example, so a regression in the TER math fails fast.

Related tools

LiveAI

NIST Score Calculator

Calculate the NIST machine-translation score for a candidate translation against one or more references, entirely in your browser. Shows the information-weighted precision per n-gram order, the NIST brevity penalty, and every matched n-gram's weight. Matches NLTK sentence_nist, no signup.

Open tool

LiveAI

METEOR Score Calculator

Calculate the METEOR score for a candidate translation against a reference, entirely in your browser. Shows unigram matches, precision, recall, the recall-weighted Fmean, the chunk-based fragmentation penalty, and the aligned tokens. Matches NLTK single_meteor_score, no signup.

Open tool

LiveAI

BLEU Score Calculator

Calculate the BLEU score for a candidate translation against one or more references, entirely in your browser. Shows the modified n-gram precisions p1–p4, the brevity penalty, and the clipped match counts. Matches NLTK sentence_bleu, no signup.

Open tool

Rate this tool

Be the first to rate

Comments & feedback

Spotted a bug or want an improvement? Tell us — our team reviews every comment, and good ideas get built. Comments are public and anonymous.

Found a bug, edge case, or want to suggest an improvement?

Email me at [email protected] — most fixes ship within 24 hours.