ucrelnlp/Multilingual-USAS-Labelled-Silver-Wikipedia

Multilingual USAS Silver Labelled Wikipedia Articles

Silver-labelled Wikipedia article text for training USAS semantic taggers and Multi-Word Expression (MWE) identifiers, covering 8 Wikipedia language sites. The source text comes from the HuggingFace HuggingFaceFW/finewiki dataset, restricted to articles rated Good (GA) or Featured (FA) — using the article ID list from ucrelnlp/wikipedia-ga-fa-ids — and then sentence split and automatically tagged with USAS semantic tags and MWEs using PyMUSAS rule-based taggers. For more information on how the dataset was generated, including the full filtering/processing pipeline, see https://github.com/UCREL/wikipedia-USAS-processing.

• Curated by: University Centre for Computer Corpus Research on Language (UCREL) group at Lancaster University
• Multi-lingual
• Repository: https://github.com/UCREL/wikipedia-USAS-processing

Uses

It can be used to train USAS semantic taggers and MWE identifiers.

Filtering and Processing

Each Wikipedia article goes through the following pipeline before being included in this dataset:

• The article ID and title must match an article rated as Good (GA) or Featured (FA) (taken from ucrelnlp/wikipedia-ga-fa-ids).
• Articles that are part of a manually curated test set (by URL) are excluded.
• Wikipedia family-tree tables, mathematical equations, and other tables are removed from the article text.
• Markdown formatting (e.g. headers like # but not the header text) is stripped from the article text.
• Articles with fewer than 50 tokens, based on a language-specific tokenizer, are removed.
• Exact and then MinHash de-duplication is applied.
• Remaining articles are sentence split using language-specific spaCy sentence splitters.
• Each sentence is tagged with USAS semantic tags and, where the tagger supports it, MWEs, using PyMUSAS Rule Based language-specific taggers.

Dataset Structure

Each row is a single article, unique per id/page_id within a language config. The data is split per language into train and validation subsets (see below).

• text - the processed article text.
• id - unique identifier for the article, e.g. enwiki/23146210
• page_id - the Wikipedia page ID 23146210
• title - Article title.
• url - the article URL, e.g. https://en.wikipedia.org/wiki/Bill_Gutteron
• version- (integer) revision/version identifier of the page (comes from HuggingFaceFW finewiki) 1230438345
• start_end_sentence_character_indexes - list of [start, end] character offsets for each sentence, e.g. [[0, 10], [11, 15]] the first sentence is between text[0:10].
• tokens - list of a list of tokens whereby the inner list represents the tokens for a given sentence, e.g. tokens[0] would contain all of the tokens in the first sentence.
• tags - list of a list of a list of USAS tags that were predicted by the PyMUSAS Rule Based languages specific tagger. The inner list represents the most likely USAS tags for the given token, e.g. tags[0][0] will contain a list of most likely USAS tags for the first token in the first sentence, in most cases it will only contain one USAS tag. When it contains more than one USAS tag this represents a token in which the meaning is a combination of the given predicted USAS tags. Some tokens will contain no USAS tags as the Rule Based tagger cannot make prediction for all tokens. Tags within each group are de-duplicated.
• other_tags - list of a list of a list of a list of USAS tags, one level deeper than tags, containing every other valid USAS tag group for a token that was not its most likely tag group, e.g. other_tags[0][0] will contain a list of the other valid USAS tag groups for the first token in the first sentence, and other_tags[0][0][0] the tags within the first of those groups. Keeping each group as its own inner list (rather than merging them together) preserves which tags PyMUSAS considered part of the same combined meaning. Most tokens will contain no other USAS tag groups, in which case the outer list is empty. As with tags, tags within each group are de-duplicated. They are ordered by the most likely USAS tag group.
• mwes - list of a list of MWE labels that were predicted by the PyMUSAS Rule Based languages specific tagger, these always relate to the most likely USAS tags. The MWE labels denote at the sentence level which tokens are MWEs, e.g. mwes[0][0] represent all of the MWE labels for the first token in the first sentence, if it contains 1 and mwes[0][1] also contains 1 then the first token and second token in the first sentence are a MWE. If more than one label occurs then MWEs are overlapping which should not be the case with PyMUSAS taggers. MWEs can be dis-continuous. The index of MWE labels always start at 1 and reset per sentence, e.g. the first sentence can contain a MWE label of 1 and so can the second sentence, but they will be different MWEs as MWEs are constrained to occur within a single sentence; they cannot span sentence boundaries.

The data is stored as zstd-compressed Parquet files.

Example of a record (shown as JSON for readability), taken directly from the English config of this dataset (enwiki/23146210, the full Bill Gutteron article). Note that text reflects the article revision captured in the underlying FineWiki snapshot (June 2024 version of the article) — the live Wikipedia article has since been substantially expanded, so it is now much longer than the short, complete example shown here:

{
  "text": "Bill Gutteron\n\nWilliam Alexander Gutteron (November 26, 1899 – May 30, 1987) was a professional football player in the National Football League (NFL). He made his NFL debut in 1926 with the Los Angeles Buccaneers. He played only one season in the league. A quarterback, Gutteron played college football for the Nevada Wolf Pack.\n",
  "id": "enwiki/23146210",
  "page_id": 23146210,
  "title": "Bill Gutteron",
  "url": "https://en.wikipedia.org/wiki/Bill_Gutteron",
  "version": 1230438345,
  "start_end_sentence_character_indexes": [[0, 150], [151, 213], [214, 254], [255, 329]],
  "tokens": [["Bill", "Gutteron", "\n\n", "William", "Alexander", "Gutteron", "(", "November", "26", ",", "1899", "–", "May", "30", ",", "1987", ")", "was", "a", "professional", "football", "player", "in", "the", "National", "Football", "League", "(", "NFL", ")", "."], ["He", "made", "his", "NFL", "debut", "in", "1926", "with", "the", "Los", "Angeles", "Buccaneers", "."], ["He", "played", "only", "one", "season", "in", "the", "league", "."], ["A", "quarterback", ",", "Gutteron", "played", "college", "football", "for", "the", "Nevada", "Wolf", "Pack", "."]],
  "tags": [[["Z1"], ["Z1"], [], ["Z1"], ["Z1"], ["Z1"], ["Z9"], ["Z2"], ["Z2"], ["Z9"], ["N1"], ["Z9"], ["Z2"], ["Z2"], ["Z9"], ["N1"], ["Z9"], ["A3"], ["Z5"], ["I3.2"], ["K5.1", "S2"], ["K5.1", "S2"], ["Z5"], ["Z5"], ["Z1"], ["Z1"], ["Z1"], ["Z9"], ["G3"], ["Z9"], ["Z9"]], [["Z8"], ["A5.4"], ["A5.4"], ["T1.3"], ["T1.3"], ["Z5"], ["N1"], ["Z5"], ["Z5"], ["Z1"], ["Z1"], ["Z1"], ["Z9"]], [["Z8"], ["K1"], ["A14"], ["N1"], ["T1.3"], ["Z5"], ["Z5"], ["S5"], ["Z9"]], [["Z5"], [], ["Z9"], [], ["K1"], ["P1", "H1"], ["K5.2"], ["Z5"], ["Z5"], ["Z1"], ["Z1"], ["Z1"], ["Z9"]]],
  "other_tags": [[[["Z3"]], [["Z3"]], [], [["Z3"]], [["Z3"]], [["Z3"]], [], [["Z1"], ["T1.3"]], [["Z1"], ["T1.3"]], [], [], [], [["Z1"], ["T1.3"]], [["Z1"], ["T1.3"]], [], [], [], [["Z5"]], [], [["I3.1"], ["X9.1"], ["A5.1"]], [], [], [], [], [["Z3"]], [["Z3"]], [["Z3"]], [], [], [], []], [[], [], [], [], [], [], [], [], [], [["Z3"]], [["Z3"]], [["Z3"]], []], [[], [["K5.1"], ["K5.2"], ["K2"], ["K3"], ["A1.1.1"], ["K6"]], [], [["T3"], ["T1.2"]], [], [], [], [["K5.1", "S5"], ["S7.3"], ["N3.3"]], []], [[], [], [], [], [["K5.1"], ["K5.2"], ["K2"], ["K3"], ["A1.1.1"], ["K6"]], [], [], [], [], [["Z3"]], [["Z3"]], [["Z3"]], []]],
  "mwes": [[[1], [1], [], [2], [2], [2], [], [3], [3], [], [], [], [4], [4], [], [], [], [], [], [], [5], [5], [], [], [6], [6], [6], [], [], [], []], [[], [1], [1], [2], [2], [], [], [], [], [3], [3], [3], []], [[], [], [], [], [], [], [], [], []], [[], [], [], [], [], [], [], [], [], [1], [1], [1], []]]
}

The first sentence contains six MWEs: ["Bill", "Gutteron"] (label 1), ["William", "Alexander", "Gutteron"] (label 2), ["November", "26"] (label 3), ["May", "30"] (label 4), ["football", "player"] (label 5), and ["National", "Football", "League"] (label 6). Labels reset per sentence, so the second sentence starts again from 1 with its own three MWEs: ["made", "his"] (label 1), ["NFL", "debut"] (label 2), and ["Los", "Angeles", "Buccaneers"] (label 3).

Train/validation split

Each language's documents are split into train and validation subsets, written to separate train/validation subfolders. The validation split is capped at whichever is reached first: a percentage of the language's documents (10%), or a fixed maximum number of documents (20 documents) — this keeps under-resourced languages (some have as few as ~200 articles) at a sensible percentage-based split, while bounding well-resourced languages' validation set to a sane absolute size. The split assignment is deterministic (hashed from each document's page ID), so re-running the pipeline reproduces the same split.

Dataset Statistics

Statistics for the full dataset (every language config combined), generated using the dataset_statistics.py, token_count_distribution.py, and usas_tag_distribution.py scripts from the processing GitHub repository. See the repositories README and the journal paper for more information about these tables.

Overview

Language	Articles	Sentences (M)	Tokens (M)	Labelled Tokens (M)	Labels per Labelled Token	Multi Tag Membership (%)	Unique Tags	MWEs (M)	MWE Tokens (%)
Chinese	2,807	0.677	15.613	10.111	4.48	21.13	215	0.048	0.62
Danish	187	0.068	1.416	1.070	1.69	13.41	213	0.036	6.10
Dutch	378	0.158	2.759	1.853	2.37	10.90	211	0.000	0.00
English	49,218	7.138	182.840	170.129	1.81	11.10	217	13.793	17.57
Finnish	865	0.221	3.428	2.507	1.85	16.14	209	0.000	0.00
Italian	1,161	0.335	9.640	7.904	2.04	12.01	219	0.096	2.13
Portuguese	3,469	0.764	17.886	13.875	2.78	15.24	218	0.133	1.61
Spanish	4,581	0.920	30.150	24.164	1.89	1.03	219	0.067	0.47
Total	62,666	10.281	263.732	231.611	2.01	11.52	220	14.173	12.49

"Labelled Tokens" are tokens with at least one USAS tag; "Labels per Labelled Token" and "Multi Tag Membership (%)" count labels from both the tags and other_tags columns, since both are positive labels when training (see Dataset Structure above). "MWE Tokens (%)" is the percentage of tokens that are part of at least one Multi-Word Expression.

Per-split breakdown (train / validation / total)

Language	Split	Articles	Sentences (M)	Tokens (M)	Labelled Tokens (M)	Labels per Labelled Token	Multi Tag Membership (%)	Unique Tags	MWEs (M)	MWE Tokens (%)
Chinese	train	2,787	0.670	15.463	10.010	4.48	21.12	215	0.047	0.62
Chinese	validation	20	0.007	0.149	0.100	4.53	22.01	214	0.000	0.64
Chinese	total	2,807	0.677	15.613	10.111	4.48	21.13	215	0.048	0.62
Danish	train	168	0.062	1.278	0.968	1.69	13.42	213	0.033	6.14
Danish	validation	19	0.006	0.138	0.102	1.68	13.37	211	0.003	5.74
Danish	total	187	0.068	1.416	1.070	1.69	13.41	213	0.036	6.10
Dutch	train	358	0.149	2.599	1.743	2.37	10.87	211	0.000	0.00
Dutch	validation	20	0.009	0.160	0.110	2.39	11.38	211	0.000	0.00
Dutch	total	378	0.158	2.759	1.853	2.37	10.90	211	0.000	0.00
English	train	49,198	7.134	182.734	170.030	1.81	11.10	217	13.785	17.57
English	validation	20	0.004	0.106	0.099	1.76	11.05	212	0.008	17.65
English	total	49,218	7.138	182.840	170.129	1.81	11.10	217	13.793	17.57
Finnish	train	845	0.216	3.356	2.454	1.85	16.21	209	0.000	0.00
Finnish	validation	20	0.005	0.072	0.053	1.82	12.95	207	0.000	0.00
Finnish	total	865	0.221	3.428	2.507	1.85	16.14	209	0.000	0.00
Italian	train	1,141	0.329	9.485	7.776	2.04	11.99	219	0.094	2.13
Italian	validation	20	0.005	0.154	0.127	2.11	12.83	216	0.002	2.25
Italian	total	1,161	0.335	9.640	7.904	2.04	12.01	219	0.096	2.13
Portuguese	train	3,449	0.760	17.784	13.795	2.78	15.24	218	0.132	1.61
Portuguese	validation	20	0.004	0.102	0.080	2.74	16.67	215	0.001	1.45
Portuguese	total	3,469	0.764	17.886	13.875	2.78	15.24	218	0.133	1.61
Spanish	train	4,561	0.917	30.044	24.078	1.89	1.03	219	0.067	0.47
Spanish	validation	20	0.003	0.106	0.086	1.90	1.13	219	0.000	0.52
Spanish	total	4,581	0.920	30.150	24.164	1.89	1.03	219	0.067	0.47
Total	train	62,507	10.237	262.745	230.853	2.00	11.51	220	14.158	12.52
Total	validation	159	0.044	0.987	0.758	2.39	14.21	220	0.014	3.35
Total	total	62,666	10.281	263.732	231.611	2.01	11.52	220	14.173	12.49

Token Length Distribution

Quantiles of token counts per sentence and per article (train + validation combined), plus the unweighted average across languages (Macro Avg).

Tokens per sentence

Language	P25	P50	P75	P90	P95	P99	Max
Chinese	9.0	19.0	31.0	45.0	57.0	93.0	2,604.0
Danish	11.0	19.0	28.0	37.0	44.0	65.0	708.0
Dutch	10.0	16.0	24.0	31.0	37.0	51.4	355.0
English	15.0	23.0	32.0	43.0	51.0	76.0	49,287.0
Finnish	9.0	13.0	18.0	24.0	29.0	46.0	1,032.0
Italian	16.0	25.0	37.0	51.0	61.0	91.0	2,241.0
Portuguese	11.0	21.0	32.0	44.0	53.0	76.0	927.0
Spanish	19.0	28.0	40.0	55.0	68.0	111.0	5,629.0
Macro Avg	12.5	20.5	30.2	41.2	50.0	76.2	7,847.9

Tokens per article

Language	P25	P50	P75	P90	P95	P99	Max
Chinese	2,805.0	4,249.0	6,642.0	10,834.2	14,255.2	22,480.4	50,943.0
Danish	4,754.5	6,948.0	9,420.5	13,227.4	14,707.0	18,841.6	23,163.0
Dutch	4,484.8	6,613.5	9,342.0	12,636.8	14,737.7	19,327.8	22,395.0
English	1,661.0	2,701.0	4,640.0	7,596.0	10,019.0	15,595.6	190,775.0
Finnish	2,174.0	3,244.0	4,694.0	6,475.4	7,872.4	14,011.9	134,271.0
Italian	4,365.0	7,181.0	11,134.0	15,575.0	17,940.0	23,001.0	45,295.0
Portuguese	1,856.0	3,517.0	7,009.0	11,653.6	14,543.8	19,877.7	60,246.0
Spanish	2,735.0	4,716.0	8,336.0	13,590.0	17,705.0	29,546.2	132,178.0
Macro Avg	3,104.4	4,896.2	7,652.2	11,448.5	13,972.5	20,335.3	82,408.2

USAS Tag Distribution

Tags are counted from both the tags and other_tags columns (train + validation combined), since both are positive labels when training.

Major tag distribution (%) — percentage share of the first character of each USAS tag (e.g. A3 and A1 both count towards major tag A):

Tag	Chinese	Danish	Dutch	English	Finnish	Italian	Portuguese	Spanish	Macro Avg
Z	13.2	33.9	27.8	38.8	26.0	28.1	24.3	39.9	29.0
A	17.5	16.0	17.9	14.7	18.9	14.1	15.8	15.2	16.2
S	12.0	7.7	8.3	7.5	11.0	11.0	10.5	8.5	9.6
N	9.0	7.7	7.8	7.0	8.9	9.3	8.1	7.1	8.1
M	6.1	5.7	7.1	3.6	6.2	5.5	5.9	3.1	5.4
T	4.2	6.0	4.9	5.9	4.4	3.9	5.4	5.3	5.0
X	6.3	4.2	4.3	3.7	4.3	6.3	3.7	3.9	4.6
Q	5.3	3.6	3.2	3.4	3.1	4.4	3.9	2.6	3.7
O	4.4	2.1	4.4	2.4	3.0	3.1	4.4	2.2	3.3
G	3.6	2.3	2.5	2.1	2.1	2.2	3.4	2.4	2.6
K	2.9	1.8	1.9	2.6	2.7	1.3	2.8	1.4	2.2
I	3.0	1.6	1.7	1.8	1.7	1.6	2.4	1.6	1.9
B	2.9	1.2	1.9	1.4	1.8	2.1	2.0	1.0	1.8
E	2.3	1.4	1.4	1.1	1.1	1.1	1.2	1.1	1.3
H	2.0	1.7	1.0	1.0	1.0	1.1	1.5	0.9	1.3
F	1.4	0.6	1.0	0.6	0.5	1.4	0.9	0.4	0.9
W	0.9	0.7	0.6	0.5	1.2	0.8	0.9	0.8	0.8
L	0.9	0.6	0.8	0.7	1.0	0.7	0.6	1.0	0.8
P	1.0	0.6	0.7	0.5	0.4	0.9	1.0	1.0	0.8
C	0.7	0.3	0.5	0.4	0.4	0.7	0.5	0.3	0.5
Y	0.6	0.5	0.2	0.3	0.3	0.3	0.6	0.3	0.4

5 most frequent individual tags (%)

The USAS tag categories are described and defined within the Introduction to the USAS category system, of which this can to some degree explain the reason for these tags being in the top-5;

• Z5 - Prepositions/adverbs/conjunctions etc.
• Z9 - Punctuation
• Z8 - Pronouns
• N1 - Numbers
• S2 - People

Tag	Chinese	Danish	Dutch	English	Finnish	Italian	Portuguese	Spanish	Macro Avg
Z5	4.5	15.8	17.9	15.9	7.7	15.7	14.1	25.8	14.7
Z9	6.6	9.4	6.7	7.8	10.6	7.4	5.9	7.7	7.8
Z8	0.6	2.9	1.6	1.9	2.8	2.9	2.8	4.0	2.4
N1	2.3	2.5	1.8	2.4	2.6	2.4	2.5	3.0	2.4
S2	2.6	1.7	1.5	1.6	3.1	2.1	2.5	1.8	2.1

5 least frequent individual tags (%) — shown in scientific notation, as these percentages are usually too small for a fixed decimal place to show meaningfully:

Tag	Chinese	Danish	Dutch	English	Finnish	Italian	Portuguese	Spanish	Macro Avg
X9	4.3×10⁻⁴	0	0	2.6×10⁻⁴	1.0×10⁻³	1.6×10⁻²	0	3.0×10⁻²	6.0×10⁻³
O4	1.8×10⁻⁵	5.5×10⁻⁵	0	3.0×10⁻⁵	0	2.7×10⁻⁴	2.2×10⁻⁴	5.7×10⁻²	7.2×10⁻³
G2	0	0	0	1.1×10⁻⁵	0	6.2×10⁻⁴	1.1×10⁻⁴	8.8×10⁻²	1.1×10⁻²
Q2	3.6×10⁻⁴	1.1×10⁻⁴	0	1.3×10⁻⁵	0	5.0×10⁻⁵	5.4×10⁻⁴	9.0×10⁻²	1.1×10⁻²
A1.5	8.3×10⁻³	0	2.1×10⁻³	9.4×10⁻⁶	0	1.3×10⁻²	1.4×10⁻³	7.1×10⁻²	1.2×10⁻²

Tag frequency spread — five-number summary (raw count, with percentage in brackets) of how spread out individual tags' frequencies are within each language:

Statistic	Chinese	Danish	Dutch	English	Finnish	Italian	Portuguese	Spanish	Macro Avg
Min	8 (0.0%)	1 (0.0%)	91 (0.0%)	29 (0.0%)	47 (0.0%)	8 (0.0%)	8 (0.0%)	2,681 (0.0%)	359 (0.0%)
P25	57,126 (0.1%)	1,312 (0.1%)	4,048 (0.1%)	219,950 (0.1%)	3,759 (0.1%)	14,254 (0.1%)	38,321 (0.1%)	33,540 (0.1%)	46,539 (0.1%)
P50	126,862 (0.3%)	3,325 (0.2%)	8,365 (0.2%)	555,499 (0.2%)	8,815 (0.2%)	31,188 (0.2%)	75,596 (0.2%)	71,384 (0.2%)	110,129 (0.2%)
P75	253,922 (0.6%)	7,524 (0.4%)	19,326 (0.4%)	1,131,444 (0.4%)	19,582 (0.4%)	72,256 (0.4%)	168,709 (0.4%)	153,408 (0.3%)	228,271 (0.4%)
Max	2,975,981 (6.6%)	286,242 (15.8%)	784,221 (17.9%)	49,001,275 (15.9%)	491,750 (10.6%)	2,536,198 (15.7%)	5,432,492 (14.1%)	11,778,807 (25.8%)	9,160,871 (15.3%)

This table shows that on average the most frequent tag accounts for 15.3% of the USAS labels and the 75% least frequent USAS tag classes make up per USAS tag class at most 0.4% of the USAS labels on average.

Filtering Funnel (FineWiki → Final Dataset)

Of every FineWiki article considered per language, the vast majority are removed for not being rated Good/Featured; the remainder are then thinned further by the held-out test-set filter, minimum-word filter, and exact/MinHash de-duplication:

Language	Good/Featured	Test URL	Min words	Exact dedup	MinHash dedup	Total removed	Kept
Chinese	1,291,263	0	1,526	196	155	1,293,140	2,815 (0.217%)
Danish	291,764	0	0	7	3	291,774	187 (0.064%)
Dutch	2,072,477	0	0	5	5	2,072,487	378 (0.0182%)
English	6,562,224	4	6	1,608	1,571	6,565,413	49,242 (0.744%)
Finnish	571,963	0	0	30	42	572,035	865 (0.151%)
Italian	1,798,167	0	0	178	252	1,798,597	1,162 (0.0646%)
Portuguese	1,131,646	0	0	167	100	1,131,913	3,470 (0.306%)
Spanish	1,943,867	0	0	265	243	1,944,375	4,590 (0.236%)

A final post-hoc de-duplication, based on th Wikipedia Page ID (keeping the highest-version duplicate and rebalancing the train/validation split) creates the final published article counts shown in Overview above:

Language	Documents After Filtering	Final Articles	Dropped	Dropped (%)
Chinese	2,815	2,807	8	0.28
Danish	187	187	0	0.00
Dutch	378	378	0	0.00
English	49,242	49,218	24	0.05
Finnish	865	865	0	0.00
Italian	1,162	1,161	1	0.09
Portuguese	3,470	3,469	1	0.03
Spanish	4,590	4,581	9	0.20
Total (matched languages)	62,709	62,666	43	0.07

License

This dataset contains text from Wikipedia, licensed under Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0) and also available under GFDL. See Wikipedia’s licensing and Terms of Use: https://dumps.wikimedia.org/legal.html

We release this data under the same license; Creative Commons Attribution-ShareAlike 4.0 (CC BY-SA 4.0) and also available under GFDL.

Dataset Card Authors

• UCREL (ucrel@lancaster.ac.uk)
• Andrew Moore / apmoore1 (a.p.moore@lancaster.ac.uk / andrew.p.moore94@gmail.com)
• Paul Rayson (p.rayson@lancaster.ac.uk)

Dataset Card Contact

• UCREL (ucrel@lancaster.ac.uk)
• Andrew Moore / apmoore1 (a.p.moore@lancaster.ac.uk / andrew.p.moore94@gmail.com)
• Paul Rayson (p.rayson@lancaster.ac.uk)