Dataset Card for NLPre-PL – fairly divided version of NKJP1M

Dataset Summary

This is the official NLPre-PL dataset - a uniformly paragraph-level divided version of NKJP1M corpus – the 1-million token balanced subcorpus of the National Corpus of Polish (Narodowy Korpus Języka Polskiego)

The NLPre dataset aims at fairly dividing the paragraphs length-wise and topic-wise into train, development, and test sets. Thus, we ensure a similar number of segments distribution per paragraph and avoid the situation when paragraphs with a small (or large) number of segments are available only e.g. during test time.

We treat paragraphs as indivisible units (to ensure there is no data leakage between different dataset types). The paragraphs inherit the corresponding document's ID and type (a book, an article, etc.).

We provide two variations of the dataset, based on the fair division of paragraphs:

fair by document's ID
fair by document's type

Creation of the dataset

We investigate the distribution over the number of segments in each paragraph. Being Gaussian-like, we divide the paragraphs into 10 buckets of roughly similar size and then sample from them with respective ratios of 0.8 : 0.1 : 0.1 (corresponding to training, development, and testing subsets). This data selection technique assures a similar distribution of segment numbers per paragraph in our three subsets. We call it fair_by_name (shortly: by_name) since it is divided equitably regarding the unique IDs of the documents.

For creating our second split, we also consider the type of document a paragraph belongs to. We first group paragraphs into categories equal to the document types, and then we repeat the above-mentioned procedure per category. This provides us with a second split: fair_by_type (shortly: by_type).

Supported Tasks and Leaderboards

This resource can be mainly used for training the morphosyntactic analyzer models for Polish. It support such tasks as: lemmatization, part-of-speech recognition, dependency parsing.

Supported versions

This dataset is available for two tagsets and in 3 formats.

Tagsets:

UD
NKJP

File formats:

conllu
conll
conll with SpaceAfter token

All the available combinations can be found below:

fair_by_name + nkjp tagset + conllu format

load_dataset("nlprepl", name="by_name-nkjp-conllu")

fair_by_name + nkjp tagset + conll format

load_dataset("nlprepl", name="by_name-nkjp-conll")

fair_by_name + nkjp tagset + conll-SpaceAfter format

load_dataset("nlprepl", name="by_name-nkjp-conll_space_after")

fair_by_name + UD tagset + conllu format

load_dataset("nlprepl", name="by_name-nkjp-conllu")

fair_by_type + nkjp tagset + conllu format

load_dataset("nlprepl", name="by_type-nkjp-conllu")

fair_by_type + nkjp tagset + conll format

load_dataset("nlprepl", name="by_type-nkjp-conll")

fair_by_type + nkjp tagset + conll-SpaceAfter format

load_dataset("nlprepl", name="by_type-nkjp-conll_space_after")

fair_by_type + UD tagset + conllu format

load_dataset("nlprepl", name="by_type-nkjp-conllu")

Languages

Polish (monolingual)

Dataset Structure

Data Instances

    "sent_id": datasets.Value("string"),
                "text": datasets.Value("string"),
                "id": datasets.Value("string"),
                "tokens": datasets.Sequence(datasets.Value("string")),
                "lemmas": datasets.Sequence(datasets.Value("string")),
                "upos": datasets.Sequence(datasets.Value("string")),
                "xpos": datasets.Sequence(datasets.Value("string")),
                "feats": datasets.Sequence(datasets.Value("string")),
                "head": datasets.Sequence(datasets.Value("string")),
                "deprel": datasets.Sequence(datasets.Value("string")),
                "deps": datasets.Sequence(datasets.Value("string")),
                "misc"

{
 'sent_id': '3',
 'text': 'I zawrócił na rzekę.',
 'orig_file_sentence': '030-2-000000002#2-3',
 'id': ['1', '2', '3', '4', '5']
 'tokens': ['I', 'zawrócił', 'na', 'rzekę', '.'],
 'lemmas': ['i', 'zawrócić', 'na', 'rzeka', '.'],
 'upos': ['conj', 'praet', 'prep', 'subst', 'interp'],
 'xpos': ['con', 'praet:sg:m1:perf', 'prep:acc', 'subst:sg:acc:f', 'interp'],
 'feats': ['', 'sg|m1|perf', 'acc', 'sg|acc|f', ''],
 'head': ['0', '1', '2', '3', '1'],
 'deprel': ['root', 'conjunct', 'adjunct', 'comp', 'punct'],
 'deps': [''', '', '', '', ''],
 'misc': ['', '', '', '', '']
}

Data Fields

sent_id, text, orig_file_sentence (strings): XML identifiers of the present text (document), paragraph and sentence in NKJP. (These allow to map the data point back to the source corpus and to identify paragraphs/samples.)
id (sequence of strings): ids of the appropriate tokens.
tokens (sequence of strings): tokens of the text defined as in NKJP.
lemmas (sequence of strings): lemmas corresponding to the tokens.
upos (sequence of strings): universal part-of-speech tags corresponding to the tokens
xpos (sequence of labels): Optional language-specific (or treebank-specific) part-of-speech / morphological tag; underscore if not available.
feats (sequence of labels): List of morphological features from the universal feature inventory or from a defined language-specific extension; underscore if not available.
head (sequence of labels): Head of the current word, which is either a value of ID or zero (0).
deprel (sequence of labels): Universal dependency relation to the HEAD of the token.
deps (sequence of labels): Enhanced dependency graph in the form of a list of head-deprel pairs.
misc (sequence of labels): Any other annotation (most commonly contains SpaceAfter tag).

Data Splits

Fair_by_name

	Train	Validation	Test
sentences	69360	7669	8633
tokens	984077	109900	121907

Fair_by_type

	Train	Validation	Test
sentences	68943	7755	8964
tokens	978371	112454	125059

Licensing Information

This work is licensed under a Creative Commons Attribution 4.0 International License.