Back-Translation Similarity

Compare token sets in a source rendering and back-translation without claiming semantic equivalence.

Description

Back-translation similarity compares unique word sets rather than sentence meaning.

Back-translation similarity compares unique word sets rather than sentence meaning. Jaccard overlap measures the fraction of vocabulary shared by the texts.^[1]

Compare token sets in a source rendering and back-translation without claiming semantic equivalence.

When to use Back-Translation Similarity

Use this tool to compare token sets in a source rendering and back-translation without claiming semantic equivalence.

Prepare the text, recorded observations, aligned figures or editorial criteria requested. Confirm what each label, count, timestamp or supplied flag represents before interpreting the result.

How to use Back-Translation Similarity and understand its results

Prepare the text, recorded observations, aligned figures or editorial criteria requested. Confirm what each label, count, timestamp or supplied flag represents before interpreting the result.
Enter sourceQuote, backTranslatedQuote, minimumSimilarityPercent using the field table and calculation rules, then run the tool.
Read jaccardSimilarityPercent, sourceOnlyTokens, backTranslationOnlyTokens, meetsThreshold. The explanation below defines the units, signs, ordering and any flags or placeholder values.
If the tool reports an input issue, correct the values and run it again. Apply the stated assumptions when interpreting a successfully returned result.

Input field	Meaning	Value type; see rules below	Requirement
sourceQuote	Source-language rendering	string	Required
backTranslatedQuote	Back-translated rendering	string	Required
minimumSimilarityPercent	Minimum token similarity (%)	number; minimum: 0; maximum: 100	Required

Enter texts containing letter/number tokens and minimum 0–100. Lowercase with en-US rules, tokenize Unicode letters/numbers and deduplicate. Return similarity, sorted source-only and back-only tokens, and inclusive gate.

similarity%=100×intersectionSize/unionSize

Units, indexing, boundary conditions and exceptional cases are defined in the calculation rules above.

Output field	Meaning	Value type
jaccardSimilarityPercent	Token-set similarity (%)	number; minimum: 0; maximum: 100
sourceOnlyTokens	Source-only tokens	list of string
backTranslationOnlyTokens	Back-translation-only tokens	list of string
meetsThreshold	Meets threshold	boolean

json

{
  "input": {
    "sourceQuote": "The council approved the budget today",
    "backTranslatedQuote": "Today the council approved a budget",
    "minimumSimilarityPercent": 60
  },
  "output": {
    "jaccardSimilarityPercent": 83.33333333333334,
    "sourceOnlyTokens": [],
    "backTranslationOnlyTokens": [
      "a"
    ],
    "meetsThreshold": true
  }
}

Limitations and assumptions

Punctuation, order and repetition are ignored. Lexical overlap cannot establish translation fidelity or semantic equivalence.

Numeric inputs must be finite. Follow the calculation-specific rules for overflow, underflow and exceptional cases. Extremely large text inputs are rejected. Missing required fields, extra fields, wrong value types and violations of the calculation rules are input errors. A successfully returned false flag is an evaluated outcome; its meaning is explained above.

Alternative or Complementary methods

Use Fuzzy Name Match: Fuzzy-match two names with a caller-supplied similarity threshold. For a complementary task, use Transcription Name Match: Compare a transcribed name with a reference using edit distance.

References

Back-Translation Similarity — tool.io

Similar or alternative tools

Fuzzy Name Match
Fuzzy-match two names with a caller-supplied similarity threshold.
Transcription Name Match
Compare a transcribed name with a reference using edit distance.