如何使用Spacy的convert保留conllu文件中的段落信息？

Question

I'm trying to convert conllu files to Spacy's jsonl format. These conllu files contain paragraph information as specified in Universal Dependencies' website . The problem is that the paragraph information is not carrying over to the jasonl converted file where each paragraph contain a single sentence.

I'm running Spacy version 2.1.3 and using only the obligatory arguments from the spacy convert command , basically python -m spacy input.conllu output_dir

Here are the first few sentences from one of my conllu files (maybe they are not to specification?). For the sake of readability, I'm only pasting the first few tokens of each sentence.

# sent_id = tp2-p1-s1
# O cansaço começou a afetar os vestibulandos no terceiro dia de exame da Fuvest.
1   O   O   DET DET gender=MASCULINE|number=SINGULAR|proper=NOT_PROPER  2   DET _   _
2   cansaço cansaço NOUN    NOUN    gender=MASCULINE|number=SINGULAR|proper=NOT_PROPER  5   NSUBJ   _   _
3   começou começar VERB    VERB    aspect=PERFECTIVE|mood=INDICATIVE|number=SINGULAR|person=THIRD|proper=NOT_PROPER|tense=PAST 5   AUX _   _

# sent_id = tp2-p1-s2
# "Estou meio cheia, mesmo", afirmou a candidata a filosofia Scyla Pereira Gouveia, 19, que fez as provas de biologia e química, de ontem, no colégio Pueri Domus.
1   "   "   PUNCT   PUNCT   proper=NOT_PROPER   2   P   _   _
2   Estou   Estar   VERB    VERB    aspect=IMPERFECTIVE|mood=INDICATIVE|number=SINGULAR|person=FIRST|proper=NOT_PROPER|tense=PRESENT    0   ROOT    _   _
3   meio    meio    NOUN    NOUN    gender=MASCULINE|number=SINGULAR|proper=NOT_PROPER  2   DOBJ    _   _
4   cheia   cheio   ADJ ADJ gender=MASCULINE|number=SINGULAR|proper=NOT_PROPER  3   AMOD    _   _

# sent_id = tp2-p1-s3
# Seu namorado, Guilherme Schneider, 18, que presta engenharia, faz exame no mesmo local.
1   Seu Seu PRON    PRON    gender=MASCULINE|number=SINGULAR|person=THIRD|proper=NOT_PROPER 2   DET _   _
2   namorado    namorado    NOUN    NOUN    gender=MASCULINE|number=SINGULAR|proper=NOT_PROPER  13  NSUBJ   _   _

# newpar id = tp2-p2
# sent_id = tp2-p2-s1
# Pelo menos um dos 38.454 convocados para a segunda fase da Fuvest tem fortes motivos para não concluir hoje as provas.
1   Pelo    Pelo    ADP ADP gender=MASCULINE|number=SINGULAR|proper=NOT_PROPER  3   ADVMOD  _   _
2   menos   menos   NOUN    NOUN    gender=MASCULINE|number=SINGULAR|proper=NOT_PROPER  1   MWE _   _
3   um  um  NUM NUM gender=MASCULINE|proper=NOT_PROPER  13  NSUBJ   _   _

I expected the output of convert to be one file containing 2 lines, one for each paragraph. I'm getting 4 lines, one for each sentence.

I would really like to avoid building a converter of my own, if at all possible.

Thanks in advance

Answer 1

As it turns out, spaCy is prepared to have paragraph information, but, as of the writing of this answer, this is unused information.

For now, in training models that are supposed to learn sentencing, it's necessary to use the --n-sents option when using the converter