DFKI-SLT/science_ie

Dataset Card for ScienceIE

Dataset Summary

ScienceIE is a dataset for the SemEval task of extracting key phrases and relations between them from scientific documents. A corpus for the task was built from ScienceDirect open access publications and was available freely for participants, without the need to sign a copyright agreement. Each data instance consists of one paragraph of text, drawn from a scientific paper. Publications were provided in plain text, in addition to xml format, which included the full text of the publication as well as additional metadata. 500 paragraphs from journal articles evenly distributed among the domains Computer Science, Material Sciences and Physics were selected. The training data part of the corpus consists of 350 documents, 50 for development and 100 for testing. This is similar to the pilot task described in Section 5, for which 144 articles were used for training, 40 for development and for 100 testing.

There are three subtasks:

• Subtask (A): Identification of keyphrases
  • Given a scientific publication, the goal of this task is to identify all the keyphrases in the document.
• Subtask (B): Classification of identified keyphrases
  • In this task, each keyphrase needs to be labelled by one of three types: (i) PROCESS, (ii) TASK, and (iii) MATERIAL.
    • PROCESS: Keyphrases relating to some scientific model, algorithm or process should be labelled by PROCESS.
    • TASK: Keyphrases those denote the application, end goal, problem, task should be labelled by TASK.
    • MATERIAL: MATERIAL keyphrases identify the resources used in the paper.
• Subtask (C): Extraction of relationships between two identified keyphrases
  • Every pair of keyphrases need to be labelled by one of three types: (i) HYPONYM-OF, (ii) SYNONYM-OF, and (iii) NONE.
    • HYPONYM-OF: The relationship between two keyphrases A and B is HYPONYM-OF if semantic field of A is included within that of B. One example is Red HYPONYM-OF Color.
    • SYNONYM-OF: The relationship between two keyphrases A and B is SYNONYM-OF if they both denote the same semantic field, for example Machine Learning SYNONYM-OF ML.

Note: The default config science_ie converts the original .txt & .ann files to a dictionary format that is easier to use. For every other configuration the documents were split into sentences using spaCy, resulting in a 2388, 400, 838 split. The id consists of the document id and the example index within the document separated by an underscore, e.g. S0375960115004120_1. This should enable you to reconstruct the documents from the sentences.

Supported Tasks and Leaderboards

• Tasks: Key phrase extraction and relation extraction in scientific documents
• Leaderboards: https://competitions.codalab.org/competitions/15898

Languages

The language in the dataset is English.

Dataset Structure

Data Instances

science_ie

An example of "train" looks as follows:

{
  "id": "S221266781300018X", 
  "text": "Amodel are proposed for modeling data-centric Web services which are powered by relational databases and interact with users according to logical formulas specifying input constraints, control-flow constraints and state/output/action rules. The Linear Temporal First-Order Logic (LTL-FO) formulas over inputs, states, outputs and actions are used to express the properties to be verified.We have proven that automatic verification of LTL-FO properties of data-centric Web services under input-bounded constraints is decidable by reducing Web services to data-centric Web applications. Thus, we can verify Web service specifications using existing verifier designed for Web applications.", 
  "keyphrases": [
    {
      "id": "T1", "start": 24, "end": 58, "type": 2, "type_": "Task"
    }, 
    ..., 
    {"id": "T3", "start": 245, "end": 278, "type": 1, "type_": "Process"}, 
    {"id": "T4", "start": 280, "end": 286, "type": 1, "type_": "Process"}, 
    ...
  ], 
  "relations": [
    {"arg1": "T4", "arg2": "T3", "relation": 1, "relation_": "Synonym-of"}, 
    {"arg1": "T3", "arg2": "T4", "relation": 1, "relation_": "Synonym-of"}
  ]
}

subtask_a

An example of "train" looks as follows:

{
  "id": "S0375960115004120_1", 
  "tokens": ["Another", "remarkable", "feature", "of", "the", "quantum", "field", "treatment", "can", "be", "revealed", "from", "the", "investigation", "of", "the", "vacuum", "state", "."], 
  "tags": [0, 0, 0, 0, 0, 1, 2, 2, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0]
}

subtask_b

An example of "train" looks as follows:

{
  "id": "S0375960115004120_2", 
  "tokens": ["For", "a", "classical", "field", ",", "vacuum", "is", "realized", "by", "simply", "setting", "the", "potential", "to", "zero", "resulting", "in", "an", "unaltered", ",", "free", "evolution", "of", "the", "particle", "'s", "plane", "wave", "(", "|ψI〉=|ψIII〉=|k0", "〉", ")", "."], 
  "tags": [0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0, 0]
}

subtask_c

An example of "train" looks as follows:

{
  "id": "S0375960115004120_3",
  "tokens": ["In", "the", "quantized", "treatment", ",", "vacuum", "is", "represented", "by", "an", "initial", "Fock", "state", "|n0=0", "〉", "which", "still", "interacts", "with", "the", "particle", "and", "yields", "as", "final", "state", "|ΨIII", "〉", "behind", "the", "field", "region(19)|ΨI〉=|k0〉⊗|0〉⇒|ΨIII〉=∑n=0∞t0n|k−n〉⊗|n", "〉", "with", "a", "photon", "exchange", "probability(20)P0,n=|t0n|2=1n!e−Λ2Λ2n", "The", "particle", "thus", "transfers", "energy", "to", "the", "vacuum", "field", "leading", "to", "a", "Poissonian", "distributed", "final", "photon", "number", "."],
  "tags": [[0, 0, ...], [0, 0, ...], ...]
}

Note: The tag sequence consists of vectors for each token, that encode what the relationship between that token and every other token in the sequence is for the first token in each key phrase.

ner

An example of "train" looks as follows:

{
  "id": "S0375960115004120_4", 
  "tokens": ["Let", "'s", "consider", ",", "for", "example", ",", "a", "superconducting", "resonant", "circuit", "as", "source", "of", "the", "field", "."], 
  "tags": [0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 2, 0, 0, 0, 0, 0, 0]
}

re

An example of "train" looks as follows:

{
  "id": "S0375960115004120_5", 
  "tokens": ["In", "the", "quantized", "treatment", ",", "vacuum", "is", "represented", "by", "an", "initial", "Fock", "state", "|n0=0", "〉", "which", "still", "interacts", "with", "the", "particle", "and", "yields", "as", "final", "state", "|ΨIII", "〉", "behind", "the", "field", "region(19)|ΨI〉=|k0〉⊗|0〉⇒|ΨIII〉=∑n=0∞t0n|k−n〉⊗|n", "〉", "with", "a", "photon", "exchange", "probability(20)P0,n=|t0n|2=1n!e−Λ2Λ2n", "The", "particle", "thus", "transfers", "energy", "to", "the", "vacuum", "field", "leading", "to", "a", "Poissonian", "distributed", "final", "photon", "number", "."], 
  "arg1_start": 2, 
  "arg1_end": 4, 
  "arg1_type": "Task", 
  "arg2_start": 5, 
  "arg2_end": 6, 
  "arg2_type": "Material", 
  "relation": 0
}

Data Fields

science_ie

• id: the instance id of this document, a string feature.
• text: the text of this document, a string feature.
• keyphrases: the list of keyphrases of this document, a list of dict.
  • id: the instance id of this keyphrase, a string feature.
  • start: the character offset start of this keyphrase, an int feature.
  • end: the character offset end of this keyphrase, exclusive, an int feature.
  • type: the key phrase type of this keyphrase, a classification label.
  • type_: the key phrase type of this keyphrase, a string feature.
• relations: the list of relations of this document, a list of dict.
  • arg1: the instance id of the first keyphrase, a string feature.
  • arg2: the instance id of the second keyphrase, a string feature.
  • relation: the relation label of this instance, a classification label.
  • relation_: the relation label of this instance, a string feature.

Keyphrase types:

{"O": 0, "Material": 1, "Process": 2, "Task": 3}

Relation types:

{"O": 0, "Synonym-of": 1, "Hyponym-of": 2}

subtask_a

• id: the instance id of this sentence, a string feature.
• tokens: the list of tokens of this sentence, obtained with spaCy, a list of string features.
• tags: the list of tags of this sentence marking a token as being outside, at the beginning, or inside a key phrase, a list of classification labels.

{"O": 0, "B": 1, "I": 2}

subtask_b

• id: the instance id of this sentence, a string feature.
• tokens: the list of tokens of this sentence, obtained with spaCy, a list of string features.
• tags: the list of tags of this sentence marking a token as being outside a key phrase, or being part of a material, process or task, a list of classification labels.

{"O": 0, "M": 1, "P": 2, "T": 3}

subtask_c

• id: the instance id of this sentence, a string feature.
• tokens: the list of tokens of this sentence, obtained with spaCy, a list of string features.
• tags: a vector for each token, that encodes what the relationship between that token and every other token in the sequence is for the first token in each key phrase, a list of a list of a classification label.

{"O": 0, "S": 1, "H": 2}

ner

• id: the instance id of this sentence, a string feature.
• tokens: the list of tokens of this sentence, obtained with spaCy, a list of string features.
• tags: the list of ner tags of this sentence, a list of classification labels.

{"O": 0, "B-Material": 1, "I-Material": 2, "B-Process": 3, "I-Process": 4, "B-Task": 5, "I-Task": 6}

re

• id: the instance id of this sentence, a string feature.
• token: the list of tokens of this sentence, obtained with spaCy, a list of string features.
• arg1_start: the 0-based index of the start token of the relation arg1 mention, an ìnt feature.
• arg1_end: the 0-based index of the end token of the relation arg1 mention, exclusive, an ìnt feature.
• arg1_type: the key phrase type of the end token of the relation arg1 mention, a string feature.
• arg2_start: the 0-based index of the start token of the relation arg2 mention, an ìnt feature.
• arg2_end: the 0-based index of the end token of the relation arg2 mention, exclusive, an ìnt feature.
• arg2_type: the key phrase type of the relation arg2 mention, a string feature.
• relation: the relation label of this instance, a classification label.

{"O": 0, "Synonym-of": 1, "Hyponym-of": 2}

Data Splits

	Train	Dev	Test
science_ie	350	50	100
subtask_a	2388	400	838
subtask_b	2388	400	838
subtask_c	2388	400	838
ner	2388	400	838
re	24558	4838	6618

Dataset Creation

Curation Rationale

More Information Needed

Source Data

Initial Data Collection and Normalization

More Information Needed

Who are the source language producers?

More Information Needed

Annotations

Annotation process

More Information Needed

Who are the annotators?

More Information Needed

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

More Information Needed

Citation Information

@article{DBLP:journals/corr/AugensteinDRVM17,
  author    = {Isabelle Augenstein and
               Mrinal Das and
               Sebastian Riedel and
               Lakshmi Vikraman and
               Andrew McCallum},
  title     = {SemEval 2017 Task 10: ScienceIE - Extracting Keyphrases and Relations
               from Scientific Publications},
  journal   = {CoRR},
  volume    = {abs/1704.02853},
  year      = {2017},
  url       = {http://arxiv.org/abs/1704.02853},
  eprinttype = {arXiv},
  eprint    = {1704.02853},
  timestamp = {Mon, 13 Aug 2018 16:46:36 +0200},
  biburl    = {https://dblp.org/rec/journals/corr/AugensteinDRVM17.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}

Contributions

Thanks to @phucdev for adding this dataset.