Dataset Card for NYT-multi

Dataset Summary

The original NYT dataset by Riedel et al. (2010) consists of New York Times news articles from 1987-2007 that was distantly annotated with relations using FreeBase. The original dataset consisted of 1.18M sentences. It is available here: https://iesl.cs.umass.edu/riedel/ecml/

The version here is mainly based on the work by Zeng et al. (2018) that filtered out sentences with more than 100 words and sentences without an active relation, leaving 66195 sentences. They randomly selected 5000 sentences from it as the test set, 5000 sentences as the validation set and the rest 56195 sentences are used as train set. The resulting dataset called NYT-multi features overlapping entities across three entity types and 24 relation types. The data is available here: https://github.com/xiangrongzeng/copy_re

The data was further pre-processed with the StanfordCoreNLP by Yu et al. (2020): https://github.com/yubowen-ph/JointER

We converted the data into a more readable JSON format. You can access the raw version from Zeng et al. (2018) using datasets.load_dataset("DFKI-SLT/nyt-multi", config="raw").

Languages

The language in the dataset is English.

Dataset Structure

Dataset Instances

default

An example of 'train' looks as follows:

{
  "tokens": ["Massachusetts", "ASTON", "MAGNA", "Great", "Barrington", ";", "also", "at", "Bard", "College", ",", "Annandale-on-Hudson", ",", "N.Y.", ",", "July", "1-Aug", "."],
  "spo_list": [["Annandale-on-Hudson", "/location/location/contains", "Bard College"]],
  "pos_tags": ["NNP", "NNP", "NNP", "NNP", "NNP", ":", "RB", "IN", "NNP", "NNP", ",", "NNP", ",", "NNP", ",", "NNP", "NNP", "."],
  "relations": [
    {
      "h": {"text": "Annandale-on-Hudson", "start": 11, "end": 12, "type": "LOCATION"},
      "t": {"text": "Bard College", "start": 8, "end": 10, "type": "ORGANIZATION"},
      "type": "/location/location/contains"
    }
  ]
}

raw

An example of 'train' looks as follows:

{
  "sentText": "Massachusetts ASTON MAGNA Great Barrington ; also at Bard College , Annandale-on-Hudson , N.Y. , July 1-Aug .",
  "articleId": "/m/vinci8/data1/riedel/projects/relation/kb/nyt1/docstore/nyt-2005-2006.backup/1669365.xml.pb",
  "relationMentions": [
    {"em1Text": "Annandale-on-Hudson", "em2Text": "Bard College", "label": "/location/location/contains"}
  ],
  "entityMentions": [
    {"start": 1, "label": "ORGANIZATION", "text": "Bard College"},
    {"start": 2, "label": "LOCATION", "text": "Annandale-on-Hudson"}
  ],
  "sentId": "1"
}

Data Fields

default

• tokens: the tokenized text of this example, a list of string features.
• spo_list: the relation triplets (head entity text, relation type, tail entity text), a list of lists containing string features.
• pos_tags: the part-of-speech tags of this example, a list of string features.
• relations: list of relations
  • h: the head entity
    • text: the entity text, a string feature.
    • start: start index of the head entity, a int32 feature.
    • end: end index of the head entity, a int32 feature.
    • type: the entity type, a string feature.
  • t: the tail entity
    • text: the entity text, a string feature.
    • start: start index of the tail entity, a int32 feature.
    • end: end index of the tail entity, a int32 feature.
    • type: the entity type, a string feature.
  • type: relation type, a string feature.

raw

• sentText: the text of this example, a string feature.
• articleId: the id of the article, a string feature.
• relationMentions: list of relation mentions
  • em1Text: the head entity text, a string feature.
  • em2Text: the tail entity text, a string feature.
  • label: relation type, a string feature.
• entityMentions: list of entity mentions
  • start: start index of the tail entity, a int32 feature.
  • label: the entity type, a string feature.
  • text: the entity text, a string feature.
• sentId: index of the sentence, a string feature

Citation

BibTeX:

@inproceedings{zeng-etal-2018-extracting,
    title = "Extracting Relational Facts by an End-to-End Neural Model with Copy Mechanism",
    author = "Zeng, Xiangrong  and
      Zeng, Daojian  and
      He, Shizhu  and
      Liu, Kang  and
      Zhao, Jun",
    editor = "Gurevych, Iryna  and
      Miyao, Yusuke",
    booktitle = "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2018",
    address = "Melbourne, Australia",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/P18-1047",
    doi = "10.18653/v1/P18-1047",
    pages = "506--514",
    abstract = "The relational facts in sentences are often complicated. Different relational triplets may have overlaps in a sentence. We divided the sentences into three types according to triplet overlap degree, including Normal, EntityPairOverlap and SingleEntiyOverlap. Existing methods mainly focus on Normal class and fail to extract relational triplets precisely. In this paper, we propose an end-to-end model based on sequence-to-sequence learning with copy mechanism, which can jointly extract relational facts from sentences of any of these classes. We adopt two different strategies in decoding process: employing only one united decoder or applying multiple separated decoders. We test our models in two public datasets and our model outperform the baseline method significantly.",
}
@article{yu-etal-2019-joint,
  author       = {Bowen Yu and
                  Zhenyu Zhang and
                  Jianlin Su and
                  Yubin Wang and
                  Tingwen Liu and
                  Bin Wang and
                  Sujian Li},
  title        = {Joint Extraction of Entities and Relations Based on a Novel Decomposition
                  Strategy},
  journal      = {CoRR},
  volume       = {abs/1909.04273},
  year         = {2019},
  url          = {http://arxiv.org/abs/1909.04273},
  eprinttype    = {arXiv},
  eprint       = {1909.04273},
  timestamp    = {Mon, 24 Aug 2020 08:57:29 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-1909-04273.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}
@inproceedings{riedel2010modeling,
  title={Modeling relations and their mentions without labeled text},
  author={Riedel, Sebastian and Yao, Limin and McCallum, Andrew},
  booktitle={Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2010, Barcelona, Spain, September 20-24, 2010, Proceedings, Part III 21},
  pages={148--163},
  year={2010},
  organization={Springer}
}

APA:

• Zeng, X., Zeng, D., He, S., Liu, K., & Zhao, J. (2018, July). Extracting relational facts by an end-to-end neural model with copy mechanism. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 506-514).
• Yu, B., Zhang, Z., Su, J., Wang, Y., Liu, T., Wang, B., & Li, S. (2019). Joint extraction of entities and relations based on a novel decomposition strategy. CoRR, abs/1909.04273. Retrieved from http://arxiv.org/abs/1909.04273
• Riedel, S., Yao, L., & McCallum, A. (2010). Modeling relations and their mentions without labeled text. In Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2010, Barcelona, Spain, September 20-24, 2010, Proceedings, Part III 21 (pp. 148-163). Springer Berlin Heidelberg.

Dataset Card Authors

@phucdev