UzSyn - A Syntactically Annotated Uzbek Dataset

UzSyn contains Uzbek sentences with grammatical and syntactic annotations at the sentence and token levels. It provides sentence structure and communicative types, part-of-speech tags, sentence constituent roles, BIO chunks, dependency relations, and bracketed constituency trees.

Repository: uznlp-uz/uzsyn

Intended Uses

• Part-of-speech tagging for Uzbek.
• Sentence constituent labeling and syntactic chunking.
• Dependency parsing.
• Constituency parsing.
• Sentence classification by structure and communicative purpose.
• Uzbek syntax research and NLP experiments.

Dataset Statistics

The following statistics were computed from the updated Sentences.tsv, Tokens.tsv, and TagSets.tsv files. Token counts include punctuation.

Statistic	Value
Sentences	1,123
Tokens	4,955
Tokens labeled XPOS = PUNCT	1,225
Non-punctuation tokens	3,730
Tokens per sentence	2-27
Sentences with a populated ConstituencyTree field	1,123
Sentences with an empty ConstituencyTree field	0
Observed XPOS tags	13
Observed sentence constituent role tags	8
Observed BIO tags	13
Observed dependency relation labels	24

Sentence Type Distribution

Field	Tag	Meaning	Sentences
SentenceType	SODDA	Simple sentence	1,098
SentenceType	QOSHMA	Compound or complex sentence	25
CommunicativeType	DARAK	Declarative sentence	951
CommunicativeType	SOROQ	Interrogative sentence	101
CommunicativeType	BUYRUQ	Imperative sentence	57
CommunicativeType	ISTAK	Optative sentence (expressing a wish)	14

Files and Schema

All files use UTF-8 encoding and tab-separated (\t) fields.

Sentences.tsv

Each data row represents one sentence. The first row contains column names.

Column	Description
sent_id	Sentence identifier; the join key for Tokens.tsv.
SentenceText	Sentence text.
SentenceType	Structural type: SODDA or QOSHMA.
CommunicativeType	Communicative type: DARAK, SOROQ, BUYRUQ, or ISTAK.
ConstituencyTree	Constituency tree represented using nested brackets.

Tokens.tsv

Each data row represents one token. The first row contains column names.

Column	Description
sent_id	Identifier of the sentence containing the token.
TokenId	Token identifier within its sentence.
token	Token text: a word or punctuation mark.
XPOS	Part-of-speech tag from the project's tagset.
GapBolagi	The token's sentence constituent role.
GapChunk	BIO label for the token's syntactic chunk.
HEAD	TokenId of the token's syntactic head within the same sentence; 0 indicates the root.
DEPREL	Dependency relation between the token and its head.

Use the pair (sent_id, TokenId) to identify a token. Do not treat sent_id as a row number or assume that sentence identifiers are consecutive.

TagSets.tsv

A tag reference containing Uzbek descriptions, usage notes, and selected examples. It has eight sections: SentenceType, CommunicativeType, XPOS, GapBolagi, GapChunk, DEPREL, ConstituencyType, and ConstituencyFunction.

This file contains section titles and repeated column headers. It is therefore excluded from the Hugging Face data configurations and should be read as a tag reference rather than a training table.

Annotation Tags

Part-of-Speech Tags - XPOS

Tag	Meaning	Tokens
N	Noun	1,454
VB	Verb	1,129
JJ	Adjective	328
NUM	Numeral	37
RR	Adverb	120
P	Pronoun	426
II	Adposition (postposition)	42
C	Conjunction	37
Prt	Particle	29
MD	Modal word	21
IM	Imitative word / ideophone	1
UH	Interjection	106
PUNCT	Punctuation	1,225

Tags are case-sensitive; for example, preserve the spelling Prt.

Sentence Constituent Roles - GapBolagi

Tag	Meaning	Tokens
EGA	Subject	949
KESIM	Predicate	1,455
ANIQLOVCHI	Attributive modifier	488
TOLDIRUVCHI	Object / complement	249
HOL	Adverbial	387
KIRISH	Parenthetical or introductory expression	15
UNDALMA	Vocative	16
BOSHQA	Other / punctuation; not assigned a sentence constituent role	1,396

Syntactic Chunks - GapChunk

B- marks the beginning of a chunk, I- marks its continuation, and the standalone O label marks a token outside a chunk.

Chunk	Tags	Role
S	B-S, I-S	Subject chunk
P	B-P, I-P	Predicate chunk
O	B-O, I-O	Object / complement chunk
A	B-A, I-A	Attributive modifier chunk
AdvC	B-AdvC, I-AdvC	Adverbial chunk
KIR	B-KIR, I-KIR	Parenthetical or introductory expression chunk
UND	B-UND, I-UND	Vocative chunk

The reference defines 15 BIO tags. In the current Tokens.tsv, B-A and I-A do not occur; the remaining 13 tags are present. Distinguish the standalone O label from B-O and I-O.

Dependency Relations - DEPREL

root, nsubj, obj, obl, advmod, amod, nmod:poss, compound,
aux, cop, case, conj, cc, mark, punct, discourse, vocative,
acl, xcomp, advmod:emph, nummod, advcl, ccomp, appos

Project-specific descriptions are provided in TagSets.tsv. These TSV files are not in CoNLL-U format. Similarity between label names does not establish full compliance with Universal Dependencies annotation guidelines.

Constituency Trees - ConstituencyTree

The tag reference defines the constituent types S, NP, VP, ADJP, ADVP, PP, CP, SBAR, ACL, and DISC, and the functions SBJ, OBJ, PRD, TMP, LOC, MNR, ADV, and OBL. For example, NP-SBJ denotes a noun phrase functioning as a subject.

Example

Sentence sent_id = 1:

Agar u rozi bo'lsa biz darhol boshlaymiz.

English translation: "If he/she agrees, we will start immediately."

• SentenceType: QOSHMA
• CommunicativeType: DARAK

TokenId	token	XPOS	GapBolagi	GapChunk	HEAD	DEPREL
1	Agar	C	BOSHQA	O	4	mark
2	u	P	EGA	B-S	4	nsubj
3	rozi	JJ	KESIM	B-P	4	compound
4	bo'lsa	VB	KESIM	I-P	7	advcl
5	biz	P	EGA	B-S	7	nsubj
6	darhol	RR	HOL	B-AdvC	7	advmod
7	boshlaymiz	VB	KESIM	B-P	0	root
8	.	PUNCT	BOSHQA	O	7	punct

Usage

Hugging Face Datasets

pip install datasets

Once this README is placed at the repository root, its YAML configurations allow the two tables to be loaded separately:

from datasets import load_dataset

sentences = load_dataset("uznlp-uz/uzsyn", "sentences", split="train")
tokens = load_dataset("uznlp-uz/uzsyn", "tokens", split="train")

print(sentences[0])
print(tokens[0])

train is a technical split name used to expose all records. The source files do not provide predefined training, validation, or test partitions. The sentences and tokens configurations represent the same sentences at two annotation levels.

For configuration details, see Hugging Face - Manual Configuration.

Reading Local TSV Files

This example uses only the Python standard library. Run it from the directory containing the TSV files:

import csv
from collections import defaultdict

def read_tsv(path):
    with open(path, encoding="utf-8-sig", newline="") as file:
        return list(csv.DictReader(file, delimiter="\t"))

sentences = read_tsv("Sentences.tsv")
tokens = read_tsv("Tokens.tsv")

tokens_by_sentence = defaultdict(list)
for row in tokens:
    row["TokenId"] = int(row["TokenId"])
    row["HEAD"] = int(row["HEAD"])
    tokens_by_sentence[row["sent_id"]].append(row)

for rows in tokens_by_sentence.values():
    rows.sort(key=lambda row: row["TokenId"])

sentence = sentences[0]
rows = tokens_by_sentence[sentence["sent_id"]]
print(sentence["SentenceText"])
print([row["token"] for row in rows])
print([row["XPOS"] for row in rows])

License

The license indicated for the repository is CC BY 4.0 (cc-by-4.0).

Citation and Contact

When using this dataset, acknowledge UzSyn and include the repository URL. Record the commit identifier used in your experiments to support reproducibility. The supplied files do not specify an associated publication or DOI.

Questions and annotation feedback can be submitted through the repository's Community discussions.