mneb/conll2004

CoNLL04

CoNLL04 (end-to-end relation extraction), rebuilt for MNEB from the original release with character-level offsets.

Source: Split from SpERT (https://github.com/lavis-nlp/spert); sentences checked against the original corpus (https://cogcomp.seas.upenn.edu/Data/ER/conll04.corp)

config	split	records
default	train	922
default	validation	231
default	test	288

Labels

entities: person, location, organization, other

relations: Located in, Work for, Organization based in, Live in, Kill

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

• Typed entities are included in output.entities.
• Offsets come from the source annotation; the previous version placed some mentions at the first matching string.
• Relation descriptions are adapted from Roth and Yih (2004).

Build

Built by scripts/datasets/conll2004.py in the MNEB repository.

Citation

Roth and Yih (2004), A Linear Programming Formulation for Global Inference in Natural Language Tasks. CoNLL.

License: not stated by the source; see the original release.