coref-data/korean_ecmt_detokenized

Detokenizes the ECMT dataset using the kiwipiepy library.

The script used to convert the dataset is here: https://gist.github.com/ianporada/a246ebf59696c6e16e1bc1873bc182a4

The library version used is kiwipiepy==0.20.3 / kiwipiepy_model==0.20.0

The dataset schema is as follows:

{
  # the original document filename
  "doc_id": str,

  # a list of sentences in the document
  "sentences": [
    "index": int, # the index of the sentence within the document
    "detokenized_text": str, # a single string representing the text of the sentence (detokenized using kiwipiepy)

    # a list of token positions which are tuples of the form (start, end)
    # the token at index i corresponds to characters detokenized_text[start:end]
    "token_positions": [(int, int), ...],

    # the original values of each token from the dataset
    "tokens": [{"index": int, "text": str, "xpos": str}, ...], 
  ],

  # a list of coreference chains, each chain is a list of mentions
  # each mention is a list of form [sentence_index, start_token_index, end_token_index] where token indices are inclusive indices within the given sentence
  "coref_chains": [[[int, int, int], ...], ...]
}