gamendez98/synthetic_gec_es

Synthetic data

This synthetic dataset is devided into chunk files of around 1000000 sentences each. with the structure:

[
  {
    "original_words": ["@@PADDING@@", "El", "ama", "los", "gatos", "."],
    "modified_words": ["@@PADDING@@", "El", "ama", "gatos", "."],
    "labels": ["$KEEP", "$KEEP", "$APPEND_los", "$KEEP", "$KEEP"]
  },
  
  {}
  
]

Where original_words has the sequence of tokens as they came from the raw data, modified_words are the tokens after injecting errors into the sentence and labels corresponds to the necessary corrections to go back to the original.

These corrections can be:

• $KEEP: No change is needed in this position
• $DELETE: The word corresponding to this position has to be removed
• $APPEND_<word>: append the token <word> after the position of this label
• $SPLIT_<place>: split the word at this position at the character <place>
• $REPLACE_<word>: replace the word at this position for <word>

Grammar error correction

In order to use the labels for error correction we provide an aditional repository (which was also used to create this data). This repository also provides an easy way to make the corrections described in the synthetic dataset like so:

from grammar_error_correction.grammar_error_correction import correct_all_errors

words= ["@@PADDING@@", "El", "ama", "gatos", "."]
labels= ["$KEEP", "$KEEP", "$APPEND_los", "$KEEP", "$KEEP"]

corrected_words = correct_all_errors(words, labels)

This functionality can be used to check the correctness of the dataset

from data_preparation.debug_tools import check_synthetic_dataset_correctness

check_synthetic_dataset_correctness('path/to/syn_data.json')