Cross-Lingual Mode (source only)
With source + candidate, SemScore measures semantic distance directly across languages. The multilingual-e5-small model projects both texts into a shared 384-dimensional space, so English and Japanese sentences expressing the same meaning sit close together. Useful when no reference translation exists, but absolute scores tend to be lower than monolingual comparison — cross-lingual embeddings are genuinely harder.
Reference Match Mode (reference only)
With reference + candidate, the comparison is monolingual: "how close is the candidate to the human reference in meaning?" Scores are typically higher in absolute terms than cross-lingual scores. Best when you trust the reference and want to measure convergence even if wording differs from the literal source.
Combined Mode (both provided)
With source + reference + candidate, both signals are computed and averaged 50/50 into a single score — plus the two underlying scores are shown. This is the most robust signal: a translation scoring well on both is both faithful to the source and conventionally well-phrased. A large gap between the two is diagnostic — e.g. high reference-match but low source-fidelity suggests the reference itself drifts from the source.
Document vs Sentence Mode
Document mode returns one overall score per candidate by embedding the full text as a single vector. Sentence mode expects pre-aligned input (one sentence per line, straight from SegPrep) and scores each row independently, producing a ranked table that highlights the weakest segments. Sentence mode is the right choice when the text is long, when you need to localise defects, or when you intend to feed specific rows into a downstream review tool.
Interpreting absolute numbers
Multilingual embedding scores saturate: well-matched pairs typically land in the 80–95 range; genuinely unrelated texts still score 30–50 because of shared function words and topic similarity. Treat the number as a relative signal between candidates rather than an absolute quality judgment.