Citron uses a JSON data format for training and evaluation purposes. Each file represents one document and contains an object with the following properties:
text The text of the document
quotes Quote annotations
coreference_groups Coreference annotations (optional)
The quotes property is an array where each quote is represented by an object with the following properties:
cue Cue span
sources An array of source spans
contents An array of contents spans
coreferences An array of source coreference spans (optional)
confidence Confidence score for predicted quotes (optional)
id Unique quote identifier (optional)
Each span is represented by an object with the following properties:
start Start character index
end End character index
text Span text
Example quote object:
{
"cue": {
"start": 364,
"end": 368,
"text": "said"
},
"sources": [
{
"start": 350,
"end": 363,
"text": "Humpty Dumpty"
}
],
"contents": [
{
"start": 302,
"end": 322,
"text": "\"When I use a word,\""
},
{
"start": 369,
"end": 423,
"text": "\"it means just what I choose it to mean — neither more nor less.\""
}
]
}
The optional coreference_groups property is an array of coreference groups. Each coreference group is an array of spans, defining the members of the group.
Example:
"coreference_groups": [
[
{
"start": 350,
"end": 363,
"text": "Humpty Dumpty"
},
{
"start": 432,
"end": 434,
"text": "he"
}
]
]
Copyright 2021 British Broadcasting Corporation.