Dataset Card for wikitext_linked

Dataset Summary

The WikiText language modeling dataset is a collection of over 100 million tokens extracted from the set of verified Good and Featured articles on Wikipedia. Dependency Relations, POS, NER tags are marked with trankit, entities are linked with entity-fishing, which also tags another field of NER tags. The dataset is available under the Creative Commons Attribution-ShareAlike License.

Compared to the preprocessed version of Penn Treebank (PTB), WikiText-2 is over 2 times larger and WikiText-103 is over 110 times larger. The WikiText dataset also features a far larger vocabulary and retains the original case, punctuation and numbers - all of which are removed in PTB. As it is composed of full articles, the dataset is well suited for models that can take advantage of long term dependencies.

Supported Tasks and Leaderboards

• masked-language-modeling
• named-entity-recognition
• part-of-speech
• lemmatization
• parsing
• entity-linking-classification

Languages

English.

Dataset Structure

Data Instances

wikitext2

• Size of downloaded dataset files: 27.3 MB
• Size of the generated dataset: 197.2 MB
• Total amount of disk used: 197.2 MB

An example of 'validation' looks as follows.

{
    'text': 'It is closely related to the American lobster , H. americanus .',
    'original_id': 3,
    'tok_span': [[0, 0], [0, 2], [3, 5], [6, 13], [14, 21], [22, 24], [25, 28], [29, 37], [38, 45], [46, 47], [48, 50], [51, 61], [62, 63]],
    'tok_upos': ['root', 'PRON', 'AUX', 'ADV', 'ADJ', 'ADP', 'DET', 'ADJ', 'NOUN', 'PUNCT', 'PROPN', 'PROPN', 'PUNCT'],
    'tok_xpos': ['root', 'PRP', 'VBZ', 'RB', 'JJ', 'IN', 'DT', 'JJ', 'NN', ',', 'NNP', 'NNP', '.'],
    'tok_dephead': [0, 4, 4, 4, 0, 8, 8, 8, 4, 8, 8, 10, 4],
    'tok_deprel': ['root', 'nsubj', 'cop', 'advmod', 'root', 'case', 'det', 'amod', 'obl', 'punct', 'appos', 'flat', 'punct'],
    'tok_lemma': [None, 'it', 'be', 'closely', 'related', 'to', 'the', 'american', 'lobster', ',', 'H.', 'americanus', '.'],
    'tok_ner': [None, 'O', 'O', 'O', 'O', 'O', 'O', 'S-MISC', 'O', 'O', 'O', 'O', 'O'],
    'ent_span': [[29, 45]],
    'ent_wikipedia_external_ref': ['377397'],
    'ent_ner': [None],
    'ent_domains': [['Enterprise']],
}

wikitext103

• Size of downloaded dataset files: 1.11 GB
• Size of the generated dataset: 7.82 GB
• Total amount of disk used: 7.82 GB

An example of 'train' looks as follows.

{
    'text': 'Vision for the PlayStation Portable .',
    'original_id': 3,
    'tok_span': [[0, 0], [0, 6], [7, 10], [11, 14], [15, 26], [27, 35], [36, 37]],
    'tok_upos': ['root', 'NOUN', 'ADP', 'DET', 'PROPN', 'PROPN', 'PUNCT'],
    'tok_xpos': ['root', 'NN', 'IN', 'DT', 'NNP', 'NNP', '.'],
    'tok_dephead': [0, 0, 5, 5, 5, 1, 1],
    'tok_deprel': ['root', 'root', 'case', 'det', 'compound', 'nmod', 'punct'],
    'tok_lemma': [None, 'vision', 'for', 'the', 'PlayStation', 'Portable', '.'],
    'tok_ner': [None, 'O', 'O', 'O', 'B-MISC', 'E-MISC', 'O'],
    'ent_span': [[15, 35]],
    'ent_wikipedia_external_ref': ['619009'],
    'ent_ner': [None],
    'ent_domains': [['Electronics', 'Computer_Science']]
}

Use following code to print the examples nicely:

def print_tokens_entities(example):
    text = example['text']
    print(
        "Text:\n"
        f"   {text}"
        "\nOrig-Id: "
        f"{example['original_id']}"
        "\nTokens:"
    )
    iterator = enumerate(zip(
        example["tok_span"],
        example["tok_upos"],
        example["tok_xpos"],
        example["tok_ner"],
        example["tok_dephead"],
        example["tok_deprel"],
        example["tok_lemma"],
    ))
    print(f"    Id  | {'token':12} | {'upos':8} | {'xpos':8} | {'ner':8}  | {'deph':4} | {'deprel':9} | {'lemma':12} | Id")
    print("---------------------------------------------------------------------------------------------------")
    for idx, (tok_span, upos, xpos, ner, dephead, deprel, lemma) in iterator:
        print(f"    {idx:3} | {text[tok_span[0]:tok_span[1]]:12} | {upos:8} | {xpos:8} | {str(ner):8}  | {str(dephead):4} | {deprel:9} | {str(lemma):12} | {idx}")

    iterator = list(enumerate(zip(
        example.get("ent_span", []),
        example.get("ent_wikipedia_external_ref", []),
        example.get("ent_ner", []),
        example.get("ent_domains", []),
    )))
    if len(iterator) > 0:
        print("Entities")
        print(f"    Id  | {'entity':21} | {'wiki_ref':7} | {'ner':7} | domains")
        print("--------------------------------------------------------------------")
        for idx, ((start, end), wiki_ref, ent_ner, ent_domains) in iterator:
            print(f"    {idx:3} | {text[start:end]:21} | {str(wiki_ref):7} | {str(ent_ner):7} | {ent_domains}")

Data Fields

The data fields are the same among all splits.

• text: string feature.
• original_id: int feature. Mapping to index within original wikitext dataset.
• tok_span: sequence of (int, int) tuples. Denotes token spans (start inclusive, end exclusive) within each sentence. Note that each sentence includes an artificial root node to align dependency relations.
• tok_upos: string feature. Universal Dependency POS tag tags. Aligned with tok_span. Root node has tag "root".
• tok_xpos: string geature. XPOS POS tag. Aligned with tok_span. Root node has tag "root".
• tok_dephead: int feature. Universal Dependency Head Node. Int refers to tokens in tok_span. Root node has head 0 (itself).
• tok_deprel: Universal Dependency Relation Description. Refers to the relation between this token and head token. Aligned with tok_span. Root node has dependency relation "root" to itself.
• tok_lemma: string feature. Lemma of token. Aligend with tok_span.
• tok_ner: string feature. NER tag of token. Marked in BIOS schema (e.g. S-MISC, B-LOC, ...) Aligned with tok_span. Root node has NER tag None.
• ent_span: sequence of (int, int) tuples. Denotes entities found by entity-fishing (start inclusive, end exclusive).
• ent_wikipedia_external_ref: string feature. External Reference to wikipedia page. You can access the wikipedia page via the url https://en.wikipedia.org/wiki?curid=<ent_wikipedia_external_ref>. Aligend with ent_span. All entities either have this field, or the ent_ner field, but not both. An empty field is denoted by the string None. Aligned with ent_span.
• ent_ner: string feature. Denotes NER tags. An empty field is denoted by the string None. Aligned with ent_span. "ent_domains": sequence of string. Denotes domains of entity. Can be empty sequence. Aligned with ent_span.

Data Splits

name	train	validation	test
wikitext103	4076530	8607	10062
wikitext2	82649	8606	10062

Dataset Creation

Curation Rationale

[More Information Needed]

Source Data

Initial Data Collection and Normalization

https://huggingface.co/datasets/wikitext

Who are the source language producers?

[More Information Needed]

Annotations

Annotation process

1. Started with wikitext2-raw-v1 and wikitext103-raw-v1 from wikitext
2. Ran datasets through Trankit. Marked all fields starting with tok.

In this step, the texts have been split into sentences. To retain the original text sections you can accumulate over original_id (examples are in order).

3. Ran datasets through entity-fishing. Marked all fields starting with ent.

Who are the annotators?

Machines powered by DFKI.

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0)

Citation Information

Please cite the original creators of wikitext, and the great people developing trankit and entity-fishing.

@misc{merity2016pointer,
      title={Pointer Sentinel Mixture Models},
      author={Stephen Merity and Caiming Xiong and James Bradbury and Richard Socher},
      year={2016},
      eprint={1609.07843},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

@inproceedings{nguyen2021trankit,
      title={Trankit: A Light-Weight Transformer-based Toolkit for Multilingual Natural Language Processing},
      author={Nguyen, Minh Van and Lai, Viet Dac and Veyseh, Amir Pouran Ben and Nguyen, Thien Huu},
      booktitle="Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: System Demonstrations",
      year={2021}
}

@misc{entity-fishing,
    title = {entity-fishing},
    howpublished = {\\url{https://github.com/kermitt2/entity-fishing}},
    publisher = {GitHub},
    year = {2016--2022},
    archivePrefix = {swh},
    eprint = {1:dir:cb0ba3379413db12b0018b7c3af8d0d2d864139c}
}

Contributions

Thanks to @GabrielKP for adding this dataset.