README.md

metadata

pretty_name: SciER
license: gpl-3.0
language:
  - en
dataset_info:
  - config_name: in_domain
    features:
      - name: input
        dtype: string
      - name: output
        dtype: json
      - name: schema
        dtype: json
  - config_name: ood
    features:
      - name: input
        dtype: string
      - name: output
        dtype: json
      - name: schema
        dtype: json
configs:
  - config_name: in_domain
    data_files:
      - split: train
        path: data/in_domain/train.jsonl
      - split: validation
        path: data/in_domain/validation.jsonl
      - split: test
        path: data/in_domain/test.jsonl
  - config_name: ood
    data_files:
      - split: test
        path: data/ood/test.jsonl

SciER

SciER (named entity recognition and relation extraction), rebuilt for MNEB from the original release with character-level offsets.

Source: https://github.com/edzq/SciER (commit db347813, SciER/PLM/*.jsonl)

config	split	records
in_domain	train	5575
in_domain	validation	713
in_domain	test	854
ood	test	580

Labels

entities: dataset, method, task

relations: part of, subclass of, subtask of, benchmark for, trained with, evaluated with, synonym of, used for, compare with

Format

Each record has input (the text), output (gold annotations grouped by label) and schema (the label set with descriptions). Offsets are character offsets into input, half-open [start, end).

"output": {
  "entities":  {"<label>": [{"text", "start", "end"}]},
  "relations": {"<label>": [{"head": {"text", "start", "end"},
                             "tail": {"text", "start", "end"}}]},
  "events":    {"<label>": [{"trigger": {"text", "start", "end"},
                             "arguments": [{"role", "text", "start", "end"}]}]}
}

Each schema item has a description and a description_source: official when adapted from the dataset's paper or guideline, mneb when written for MNEB. Relation datasets also give setting (end_to_end or gold_entity) and negative_label; relation items give symmetric. Event items list the roles valid for that event type.

Within a record, identical annotations are kept once and outer whitespace is trimmed from spans; source annotations are otherwise kept as annotated.

Notes

• Configs in_domain (train/validation/test) and ood (test). Sentence level. Not SciERC.
• Descriptions are adapted from the SciER annotation guideline.

Build

Built by scripts/datasets/sci_er.py in the MNEB repository.

Citation

Zhang et al. (2024), SciER: An Entity and Relation Extraction Dataset for Datasets, Methods, and Tasks in Scientific Documents. EMNLP.

License: gpl-3.0