NagaYu/signpost-label-quality

Signpost label quality

Labels from accessibility trees, each one classed as good or as one of seven ways a label can fail to mean anything. It is built for the question that is left over after axe-core and Xcode's Accessibility Inspector have both passed: there is a name, but does the name identify this control?

Repository: NagaYu/signpost-label-quality

Code, evaluation, and the builder for this dataset	https://github.com/NagaYu/signpost
Model trained on it	https://huggingface.co/NagaYu/signpost-classifier
Try it in your browser	https://huggingface.co/spaces/NagaYu/signpost

Configs

config	train	test	total	what a row is
labels	1648	827	2475	one label, one class from the eight-way taxonomy
rewrites	622	309	931	a corrupted label and the original it replaced
hard	205	96	301	a defect that every syntactic check still passes

labels is the default config.

Data statement -- read this before quoting a number from it

The corpus is synthetic. Every screen in this build was composed by scripts/make_seed_corpus.py from a per-language vocabulary written for the project. No site was crawled and no app was decompiled, which is why the whole pipeline runs offline on a laptop and why there is no redistribution question to answer. Every row carries synthetic: true.

What that means for the numbers: a model evaluated on this data is being evaluated against a generator, and its score is partly a measurement of how learnable that generator is. Real screens are messier -- more abbreviations, more house style, more labels that are bad in ways nobody thought to plant. Take detection and accuracy figures computed here as an upper bound on a real codebase, not as an estimate of one. signpost/collect.py is the path for real pages, and it keeps the same redistribution discipline: label, role, minimal context, never a page body.

How the defects were planted. The clean corpus goes through signpost/inject.py, which corrupts a fixed share of labels (35% here) and records, for each one, the position, the class, the original string, and the strategy that produced it. Seven classes are planted: generic, role_echo, identifier, duplicate, verbose, missing and mismatch. A corrupted label is admitted to the hard set only if the corrupted string still passes every syntactic criterion -- specific, role-free, not an identifier, short enough -- so "syntactically valid but contextually wrong" is enforced by re-running the checks rather than by assertion. mismatch is generated five ways; one of them, confusable_action, draws from the same action lexicon a rule-based checker could use, and it is flagged as lexicon_derivable so nobody has to take a hard-set score at face value.

What the splits guarantee. The split is by origin -- site or app -- not by screen and not by row. Naming conventions belong to a codebase, so a per-screen split would let a model score well on a second page of a shop it had already seen. Train and test share no origin: 40 origins on the train side, 20 on the test side, checked by collect.assert_no_origin_leak before this file was written. The two sides are corrupted with different injection seeds, so a test defect is not a training defect relocated. The split is stratified by language, so no language lands entirely on one side.

What is not in the denominator. Before injection, every label passes the strict quality gate in signpost/quality.py; 0 of 2592 nodes (0.0%) were dropped as not demonstrably good. A node that was never corrupted but does not pass that gate is dropped rather than published as ok. This matters because the ok rows are the denominator of any over-flagging rate computed from this dataset, and a doubtful label counted as a good one would make every such rate -- ours included -- look better than it is.

What is not published. The label, the role, and clipped context. Page bodies, DOM, screenshots and full visible text stay local. The caps are enforced by collect.assert_no_bulk_text, which raises rather than truncates, and this build fails rather than shipping an oversized row.

Classes

class	train	test
ok	1026	518
generic	120	55
role_echo	89	48
identifier	83	54
duplicate	77	36
verbose	65	36
missing	69	27
mismatch	119	53

Languages

language	train rows	test rows
de	160	79
en	169	84
es	170	83
fr	163	86
it	164	85
ja	166	79
ko	167	83
pt	161	82
ru	164	86
zh	164	80

Mismatch strategies

mismatch strategy	train	test	a fixed lexicon can solve it
asset_export	21	12	no
asset_file	24	18	no
asset_name	10	3	no
blank_empty	49	20	no
blank_space	12	6	no
blank_zwsp	8	1	no
camelCase	7	3	no
confusable_action	38	13	yes
copy_sibling_name	77	36	no
design_tool	10	12	no
entity_swap	33	16	no
expand_to_sentence	65	36	no
generic_term	120	55	no
role_prefix	11	8	no
role_suffix	55	34	no
role_suffix_title	23	6	no
sibling_swap	25	10	no
snake_case	11	6	no
state_flip	0	1	no
topic_drift	23	13	no

Fields

field	what it holds
name	the accessible label -- the string under judgement
role	the normalised role: button, link, image, textfield, ...
value, description, visible_text	clipped to 80-120 characters
nearby_text, sibling_labels	at most 3 items of at most 80 characters
screen_title, asset_name, action_hint	the icon and the handler
lang, platform	language tag; web / ios / android
origin, url, license, license_url, attribution	provenance, per row
label	one of ok, generic, role_echo, identifier, duplicate, verbose, missing, mismatch
hard, strategy, rule_detectable	how the defect was planted
target (rewrites only)	the held-out original label
lexicon_derivable (hard only)	true for the confusable_action slice

Limitations

• Synthetic throughout. See the data statement.
• The per-row license and attribution fields are real fields carrying synthetic values: the generator assigns each origin a licence so that the provenance machinery is exercised end to end. For a corpus collected with signpost/collect.py they carry what the source actually said.
• The class balance is a choice, not an observation. Most real screens have far fewer defects than 35% of labels.
• duplicate is only definable within a screen, so a row taken out of its screen cannot be re-judged for that class.
• Nothing here says what any person will experience. The classes point at documented guidelines (WCAG 2.2, Apple's and Google's own guidance) and the citations live in signpost/cite.py. Signpost is not a replacement for axe-core, for Xcode's Accessibility Inspector, or for review by people who use assistive technology; it narrows what a human has to look at.

Reproducing this build

python scripts/make_seed_corpus.py
python scripts/build_dataset.py --seed 0 --test-frac 0.30 --inject-rate 0.35

Both are deterministic: same seed, same bytes.

Licence

The data is released under mit. Per-row licence, origin and attribution travel with each row.