sentence_spans.py

import spacy
import json
import requests
import pandas as pd

DATASET_URL = "https://raw.githubusercontent.com/liafacom/faquad/6ad978f20672bb41625b3b71fbe4a88b893d0a86/data/dataset.json"

class SentenceModel(object):
    def __init__(self, model_name):
        self.nlp = spacy.load(model_name)

    def parse(self, text):
        with self.nlp.select_pipes(enable=['tok2vec', "parser", "senter"]):
            doc = self.nlp(text)
            sentences = [ (sentence.start_char, sentence.end_char) for sentence in doc.sents ]
        return sentences

def context_generator(url):
    response = requests.get(url)
    data = json.loads(response.text)["data"]
    for idx, row in enumerate(data):
        for paragraph_idx, paragraph_row in enumerate(row["paragraphs"]):
            context = paragraph_row["context"]
            yield idx, paragraph_idx, context

def define_split(document_index):
    if document_index % 5 == 0:
        return "test"
    elif document_index % 5 == 1:
        return "validation"
    else:
        return "train"

def main():
    df = []
    model = SentenceModel("pt_core_news_sm")
    for idx, paragraph_idx, context in context_generator(DATASET_URL):
        for start_char, end_char in model.parse(context):
            row = {
                "document_index": idx,
                "paragraph_index": paragraph_idx,
                "sentence_start_char": start_char,
                "sentence_end_char": end_char,
                "split": define_split(idx)
            }
            df.append(row)
    df = pd.DataFrame(df)
    df.to_csv("spans.csv", index=False)

if __name__ == "__main__":
    main()