ryokai 了解

Ryokai (了解, "understood / got it") — a unified Python library for semantic machine-translation evaluation, combining the strengths of MEANT 2.0, XMEANT, YiSi-1/2, WOLVESAAR, and SimAlign behind one clean API on top of modern multilingual embeddings.

Pure PyTorch + HuggingFace transformers — no Stanza, no spaCy, no external parsers. Two HF models cover all 13 supported languages (en, de, fr, es, cs, fi, hi, lv, pl, ro, ru, tr, zh) in a single install:

• POS / shallow SRL: wietsedv/xlm-roberta-base-ft-udpos28 — ~1.1 GB, downloaded once.
• Multilingual embeddings: sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 — ~110 MB, used for both same-language and cross-language similarity.

Both are one-line swappable for any modern multilingual encoder (Qwen3-Embedding, Jina v3, BGE-M3, Nemotron-8B…) — see Embedding backbones in DOCUMENTATION.md.

Install

pip install ryokai

Quickstart

from ryokai import Ryokai

scorer = Ryokai()
src_lang, tgt_lang = "en", "ja"

# Most common: reference-free, word alignment + embedding
# (XMEANT-lite / YiSi-2 / Doc-embedding adequacy cross-lingual)
scorer.score(source=src, hypothesis=hyp,
             source_lang=src_lang, target_lang=tgt_lang)

Variants

One .score() call, four modes, dispatched by which arguments you pass. srl=False is the default — ryokai is no longer MEANT-first.

from ryokai import Ryokai
scorer = Ryokai()
src_lang, tgt_lang = "en", "ja"

# Reference-free, word alignment + embedding (default, most common)
# E.g. Doc-embedding adequacy / YiSi-2 / XMEANT-lite
scorer.score(source=src, hypothesis=hyp,
             source_lang=src_lang, target_lang=tgt_lang)

# Reference-based, word alignment + embedding
# E.g. Doc-embedding adequacy / WOLVESAAR / YiSi-1 / SimAlign style
scorer.score(reference=ref, hypothesis=hyp, target_lang=tgt_lang)

# Reference-free, frame-based — XMEANT proper
scorer.score(source=src, hypothesis=hyp,
             source_lang=src_lang, target_lang=tgt_lang, srl=True)

# Reference-based, frame-based — MEANT 2.0
scorer.score(reference=ref, hypothesis=hyp, target_lang=tgt_lang, srl=True)

See DOCUMENTATION.md for flags, aligner choices, embedding-backbone swaps, AER evaluation harness, CLI, architecture, and custom role weights.

References

Ryokai is glue around several published techniques — credit belongs to their authors.

License

MIT — see LICENSE.