README.md

metadata

language:
  - gl
license: cc-by-4.0
task_categories:
  - token-classification
task_ids:
  - named-entity-recognition
  - part-of-speech
pretty_name: TreebankNos
size_categories:
  - 1K<n<10K
tags:
  - galician
  - universal-dependencies
  - conllu
  - ner
  - pos
  - dependency-parsing
  - treegal
  - pud

TreebankNos

TreebankNos is a Galician-language dataset combining Universal Dependencies (UD) treebank annotations with Named Entity Recognition (NER) labels. It is built upon two established UD corpora — UD TreeGal and UD Parallel Universal Dependencies (PUD) — extended with BIO-format NER annotations covering four entity types: Person, Location, Organisation, and Miscellaneous.

The dataset is intended for multi-task NLP research in Galician, supporting POS tagging, morphological analysis, dependency parsing, and NER from a single unified resource.

---

Dataset Structure

Splits

Split	Source treebank	Sentences	Tokens
train	UD TreeGal	600	15,436
test	UD TreeGal	400	10,112
test_pud	UD PUD	1,000	23,510

Tokens are counted at the syntactic word level (multiword tokens such as contractions are expanded and each component counted separately; the surface multiword form is retained in the file but excluded from this count).

Data Format

Files follow an extended CoNLL-U format. Each token is represented as a tab-separated line with 10 fields. Sentences are separated by blank lines, and each sentence is introduced by comment lines (# sent_id and # text). Multiword tokens (e.g., Galician contractions like ó, na, do) are represented with a hyphenated ID on a dedicated line followed by their expanded components.

Column Schema

Column	Field	Description
1	ID	Token index within the sentence (integer, or range N-M for multiword tokens)
2	FORM	Surface word form
3	LEMMA	Lemma or base form
4	UPOS	Universal POS tag (UD v2 tagset)
5	XPOS	Language-specific POS tag
6	FEATS	Morphological features (key=value pairs separated by |, or _)
7	HEAD	Index of the syntactic head token (0 = root)
8	DEPREL	Universal dependency relation to the head
9	NER	NER tag in BIO format (added in TreebankNos); _ for multiword surface tokens
10	MISC	Miscellaneous annotations (e.g., SpaceAfter=No; _ if none)

Note on column 9: The NER column is not part of the original CoNLL-U standard. It is appended as the ninth column in place of the standard DEPS (enhanced dependencies) field, which is not used in these files. Multiword token lines carry _ in this column.

Example

# sent_id = 6
# text = Á dereita, outros participantes no foro, Paulo González Mariñas e Ramón Villares.
1-2	Á	_	_	_	_	_	_	_	_
1	A	a	ADP	P	AdpType=Prep	3	case	O	_
2	a	o	DET	Ddfs	Definite=Def|Gender=Fem|Number=Sing|PronType=Art	3	det	O	_
3	dereita	dereita	NOUN	Scfs	Gender=Fem|Number=Sing	6	nmod	O	SpaceAfter=No
4	,	,	PUNCT	Q,	_	3	punct	O	_
5	outros	outro	DET	Idmp	Gender=Masc|Number=Plur|PronType=Ind	6	det	O	_
6	participantes	participante	NOUN	Scmp	Gender=Masc|Number=Plur	0	root	O	_
...
11	Paulo	paulo	PROPN	Sp00	_	6	appos	B-PER	_
12	González	gonzález	PROPN	Sp00	_	11	flat:name	I-PER	_
13	Mariñas	mariñas	PROPN	Sp00	_	11	flat:name	I-PER	_
14	e	e	CCONJ	Cc	_	15	cc	O	_
15	Ramón	ramón	PROPN	Sp00	_	11	conj	B-PER	_
16	Villares	villares	PROPN	Sp00	_	15	flat:name	I-PER	SpaceAfter=No
17	.	.	PUNCT	Q.	_	6	punct	O	_

---

Label Sets

NER Tags (column 9 — BIO scheme)

Tag	Description
O	Outside any named entity
B-PER	Beginning of a Person entity
I-PER	Inside a Person entity
B-LOC	Beginning of a Location entity
I-LOC	Inside a Location entity
B-ORG	Beginning of an Organisation entity
I-ORG	Inside an Organisation entity
B-MISC	Beginning of a Miscellaneous named entity
I-MISC	Inside a Miscellaneous named entity
_	Not applicable (multiword surface token line)

NER Label Distribution

Tag	train	test (TreeGal)	test (PUD)
O	14,187	9,278	21,462
B-PER	203	141	420
I-PER	177	125	185
B-LOC	133	110	399
I-LOC	41	12	177
B-ORG	220	150	276
I-ORG	137	101	246
B-MISC	140	87	151
I-MISC	197	108	194

Universal POS Tags (column 4)

ADJ, ADP, ADV, AUX, CCONJ, DET, INTJ, NOUN, NUM, PRON, PROPN, PUNCT, SCONJ, SYM, VERB, X

Dependency Relations (column 8 — selection)

acl, advcl, advmod, amod, appos, aux, case, cc, ccomp, conj, cop, csubj, det, expl, fixed, flat:name, iobj, mark, nmod, nsubj, nummod, obj, obl, parataxis, punct, root, xcomp, and others per UD v2 guidelines.

---

Source Corpora

UD TreeGal

The TreeGal portion (train and test splits) derives from the UD_Galician-TreeGal treebank, a manually annotated corpus of Galician texts covering newspaper and political discourse. The original annotations follow the Universal Dependencies v2 framework.

UD PUD (Parallel Universal Dependencies)

The test_pud split is drawn from the UD_Galician-PUD corpus, part of the multilingual PUD collection released alongside UD v2. The PUD sentences are translations of a shared English source and include a parallel_id metadata field linking each sentence to its parallel in other languages.

---

Languages

• Language: Galician (gl)
• Script: Latin

Galician is a Romance language spoken primarily in the autonomous community of Galicia in northwestern Spain and in border regions of Portugal. It is closely related to Portuguese and shares many morphosyntactic properties with other Iberian Romance languages, including a rich inflectional system for gender, number, case (for pronouns), and verb tense/mood/aspect.

---

Intended Uses

TreebankNos is designed for:

• NER in Galician, a comparatively low-resource language for this task
• Joint / multi-task learning combining NER with POS tagging, morphological analysis, and dependency parsing
• Cross-lingual transfer studies using the PUD split, which is parallel across 20+ languages
• Benchmarking NLP pipelines for Galician

Funding

This dataset was compiled within the Nós Project, funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of the project ILENIA with reference 2022/TL22/00215336.