NagaYu/downstep-bench

Downstep compound pitch-accent benchmark

Japanese noun compounds, their mora segmentation, and every accent their source dictionaries attest. Built so that "the model has never seen this compound" is a condition you can actually turn on, rather than a claim you have to trust.

No human annotation is present in this release. Every accent label here is dictionary-derived. docs/ANNOTATION_GUIDELINES.md ships the protocol, the CSV format and the agreement statistics for collecting native-speaker judgements on novel compounds — but that collection has not happened, and nothing in this repository should be read as if it had. There are no annotator IDs, no agreement scores and no human-adjudicated items.

---

Sources and licences

Two independently produced lexicons. Downstep reads the copies installed by unidic-lite and pyopenjtalk and never fetches anything over the network.

Source	Holder	Licence	Reached via
UniDic 2.1.2 (unidic-mecab)	The UniDic Consortium / NINJAL	GPL / LGPL / BSD 3-Clause, tri-licensed — Downstep relies on the BSD option	unidic-lite
NAIST-jdic (Open JTalk's system dictionary)	Nara Institute of Science and Technology	BSD 3-Clause	pyopenjtalk

Attribution both licences require, reproduced in full in NOTICE and required of anyone redistributing these rows:

• Retain the copyright notice, the list of BSD 3-Clause conditions and the warranty disclaimer.
• Do not use "The UniDic Consortium", "NINJAL", "Nara Institute of Science and Technology" or their contributors to endorse or promote anything derived from this data without prior written permission. Downstep makes no such endorsement claim.

The BSD option is what makes the derived rows redistributable at all: under the GPL option a redistributed derivative would have to be GPL, which is why the choice is stated explicitly rather than left to the reader.

The tooling in this repository is Apache-2.0. The data rows below are governed by BSD 3-Clause through their sources.

What is redistributed, and what is not

Redistributed: derived rows — surface, reading, mora segmentation, the set of attested accents, the component decomposition, and the split/stratum labels.

Not redistributed: the dictionaries themselves, in compiled or CSV form; any audio; any human judgement. downstep.lexicon.load_lexicons() rebuilds the local cache (data/lexicon/*.jsonl, git-ignored) from the installed packages in ~20 s.

Measured in the cached dump on 2026-09-03 (make lexicon, then a count over the JSONL):

	UniDic 2.1.2	NAIST-jdic
entries	751,705	788,331
distinct surfaces	568,123	461,975
noun entries	277,189	455,743
distinct noun surfaces with an accent	258,371	418,671
noun entries recording more than one accent	21,568	0
noun surfaces whose accents disagree across entries	24,844	5,854

Those two right-hand columns are the raw material for the multiple-answer field. NAIST-jdic records one accent per entry, so its variation only appears when two entries share a surface; UniDic records variation inside a single entry ("0,4,3,1", CSV-quoted — a naive split(",") both mis-counts the fields and destroys exactly these rows).

---

What one row is

A noun compound that satisfies both of:

1. It is a lexicalised entry in NAIST-jdic, so a lexicographer assigned it an accent. That is the gold.
2. It decomposes into two or more morphemes whose readings compose back to the compound's own reading, with 連濁 / 促音化 accounted for. That is what makes it a composition problem rather than a lookup.

Cut points for (2) are constrained to a morphological analyser's boundaries. The analyser is UniDic while the gold is NAIST-jdic — deliberately a different dictionary, so the decomposition is not read off the answer. Without that constraint the decomposition collapses: a naive dynamic program over dictionary surfaces splits ファインマン into ファ+イン+マ+ン, because nearly every single kana is itself an entry, and reading-consistency alone cannot reject it.

Fields

Field	Type	Description
surface	str	The compound
reading	str	Katakana reading, devoicing marks and interpuncts removed
moras	list[str]	Mora segmentation; len(moras) == n_moras is enforced at build time
n_moras	int	Mora count
gold_accents	list[int]	A set, not a value. 0 = 平板; k = nucleus on mora k, 1-indexed
has_variation	bool	len(gold_accents) > 1 — the dictionaries attest more than one accent
components	list[obj]	Per component: surface, reading, accent, n_moras, goshu, acon_type, in_lexicon
n_components	int	Number of components (always ≥ 2)
goshu	str	語種 of the compound, taken as the modal component value (和 / 漢 / 外 / 固 / 混)
script	str	kanji / katakana / mixed
is_proper	bool	The NAIST entry's POS is 固有名詞
all_components_in_lexicon	bool	Every component was found in a source lexicon
split	str	train / dev / test
hard	bool	Test-only: no component of this compound occurs in any training compound
source	str	naist-jdic, or naist-jdic+unidic when the accent set draws on both

gold_accents is the multiple-accepted-answers field

Japanese pitch accent genuinely admits more than one correct answer for many words, and the dictionaries record that. Scoring against a single "true" accent would manufacture errors out of real variation. So:

• Score membership, not equality: a prediction is correct if it is in gold_accents. downstep.metrics.nucleus_exact_match takes a set.
• has_variation marks these rows so they can be reported as their own stratum, because a system that always guesses the majority variant looks better on them than it deserves.
• This field records lexicographic variation between and within dictionaries. It is not, and must not be presented as, a survey of how speakers pronounce these words.

---

Splits, conditions and strata — three different things

Splits partition the rows. Assignment is sha1(surface) thresholded, not a shuffled seed, so adding new candidate words never moves an existing word between splits and any machine reproduces the same partition.

split	share	rows
train	79.9%	71,544
dev	10.0% (calibration and abstention threshold)	8,959
test	10.1%	9,036

Total: 89,539 compounds.

Conditions are not splits. listed and unlisted are the same test rows evaluated twice, with the dictionary's whole-word entry present and then ablated at query time. See docs/METHODOLOGY.md for why the ablation must be rebuilt per target — batching corrupted 47.7% of targets in the pilot.

hard is a partition of the test set. Two nested flags:

• hard (818 rows, 9.0% of test) — the compound's final morpheme never occurs in any training compound. The final element (後部要素) governs compound accent in the classical rules, so this is the linguistically meaningful generalisation test.
• hard_all (105 rows) — no component occurs in training. Stricter, but too small to carry a claim; kept as a stratum.

Reported separately, never averaged into the headline: a pilot found a memorisation-shaped model gained nothing on unseen final components (58.3% vs 58.3%). The trained model does better than that — but the split exists so that can be checked rather than assumed.

stratum field	why it is a stratum
is_proper	Proper nouns are a different problem with different conventions; they are 66% of these rows (59,430 of 89,539) and are reported on their own, never folded into a single number
goshu	和語 / 漢語 / 外来語 compounds follow different accent regularities
n_components	Two-part compounds are much easier than three-part ones
n_moras	Longer compounds have more positions to get wrong
has_variation	Rows with more than one accepted answer
script	Katakana (largely loanword) compounds behave unlike kanji ones

Row counts, from benchmarks/stats.json:

stratum	counts
script	kanji 55,237 · mixed 27,290 · katakana 7,012
語種 (goshu)	和 34,760 · 固 25,085 · 漢 23,313 · 外 4,443 · 混 604 · 記号 570 · unknown 764
components	2 → 70,419 · 3 → 16,286 · 4 → 2,539 · 5+ → 295
moras	4 → 21,228 · 5 → 19,522 · 6 → 18,884 · 7 → 11,112 · 8+ → 18,793
noun type	proper 59,430 · common 30,109
more than one attested accent	1,030
every component in a lexicon	89,539 (all of them — decomposition requires it)

---

Limitations, stated plainly

• The gold is a dictionary's opinion. Dictionaries disagree with each other and with speakers. Where they disagree with each other the row keeps both answers; where they disagree with speakers, this data cannot tell.
• The variety recorded is a broadly Tokyo-based standard. That is what these two lexicons encode, so it is a scoping decision forced by the available data. It is not a claim that other regional or generational accent systems are incorrect, and a Downstep prediction is not a correction of anyone's speech.
• These are all words that exist in a dictionary. The target application is compounds that do not. The ablation device is what bridges that gap, and its limits are argued — not waved away — in docs/METHODOLOGY.md.
• Reading-consistent decomposition rejects real compounds. The rendaku model covers 連濁 and 促音化 but not 連声 or stem alternations (雨 → あま, 酒 → さか), so compounds built on those are dropped rather than mis-segmented. That is a coverage limitation the numbers do not show.
• No speech. No F0 contours, no durations, no speaker information. Accent here is an integer, not a measurement.
• Not a pronunciation dictionary. Do not use these rows as a reference for how a word "should" be said. For that, consult NHK日本語発音アクセント新辞典 or 明解日本語アクセント辞典.

Out-of-scope uses

• As an authority on correct pronunciation, in education or assessment.
• To evaluate or rank speakers, or to flag any pronunciation as an error.
• As evidence about regional variation, which it contains no information about.

Reproducing

pip install -e ".[all]"
make bench        # writes benchmarks/{train,dev,test}.jsonl and benchmarks/stats.json

scripts/build_bench.py is the only path from the two dictionaries to these rows, and it asserts its own hard property rather than trusting it. Nothing is pushed anywhere: --push exists, defaults off, and refuses to publish on its own.

Citation

Downstep: a compound pitch-accent benchmark with a dictionary-ablation
condition for words the TTS dictionary does not have.
Derived from UniDic 2.1.2 (The UniDic Consortium / NINJAL, BSD-3 option)
and NAIST-jdic (Nara Institute of Science and Technology, BSD-3).
No human annotation.

---

Substantiates: unlisted-word accuracy — these rows plus the ablation condition are what "the dictionary does not have this compound" is measured on; and reproducibility, since make bench is the only path from the two dictionaries to them.