disrpt._py

import datasets
import csv
import requests
import pandas as pd
import inspect
import copy
from .process_underscores import run

key_to_entry = requests.get('https://www.dropbox.com/scl/fi/85pnc7n6e4puoureavtzo/filtered_disrpt.json?rlkey=6cbgbe9vn2549eths7ah8gm7u&dl=1').json()
citation="\n".join(key_to_entry.values())

datasets_and_citations = {
    "deu.rst.pcc": "stede-neumann-2014-potsdam",
    "eng.dep.covdtb": "nishida-matsumoto-2022-domain",
    "eng.dep.scidtb": "yang-li-2018-scidtb",
    "eng.rst.gum": "Zeldes2017",
    "eng.rst.rstdt": "carlson-etal-2001-building",
    "eng.sdrt.stac": "asher-etal-2016-discourse",
    "eus.rst.ert": "IruskietaAranzabeIlarrazaEtAl2013",
    "fas.rst.prstc": "shahmohammadi2021persian",
    "fra.sdrt.annodis": "afantenos-etal-2012-empirical",
    "nld.rst.nldt": "redeker-etal-2012-multi",
    "por.rst.cstn": "CardosoMazieroRosarioCastroJorgeEtAl2011",
    "rus.rst.rrt": "toldova-etal-2017-rhetorical",
    "spa.rst.rststb": "da-cunha-etal-2011-development",
    "spa.rst.sctb": "cao-etal-2018-rst",
    "zho.dep.scidtb": "yi-etal-2021-unifying,cheng-li-2019-zero",
    "zho.rst.gcdt": "peng_gcdt_2022,peng_chinese_2022",
    "zho.rst.sctb": "cao-etal-2018-rst",
    "eng.pdtb.pdtb": "prasad-etal-2014-reflections",
    "eng.pdtb.tedm": "zeyrek-etal-2018-multilingual,zeyrek2019ted",
    "ita.pdtb.luna": "tonelli-etal-2010-annotation,RiccardiStepanovChowdhury2016",
    "por.pdtb.crpc": "CRPC-DB-Portuguese,genereux-etal-2012-introducing",
    "por.pdtb.tedm": "zeyrek-etal-2018-multilingual,zeyrek2019ted",
    "tha.pdtb.tdtb": "",
    "tur.pdtb.tdb": "zeyrek-webber-2008-discourse,zeyrek-kurfali-2017-tdb",
    "tur.pdtb.tedm": "zeyrek-etal-2018-multilingual,zeyrek2019ted",
    "zho.pdtb.cdtb": "Zhou2014"
}


class Config(datasets.BuilderConfig):
    citation=citation

files = [
    "deu.rst.pcc",
    "eng.dep.covdtb",
    "eng.dep.scidtb",
    "eng.pdtb.gum",
    "eng.pdtb.pdtb",
    "eng.pdtb.tedm",
    "eng.rst.gentle",
    "eng.rst.gum",
    "eng.rst.rstdt",
    "eng.sdrt.stac",
    "eus.rst.ert",
    "fas.rst.prstc",
    "fra.sdrt.annodis",
    "ita.pdtb.luna",
    "nld.rst.nldt",
    "por.pdtb.crpc",
    "por.pdtb.tedm",
    "por.rst.cstn",
    "rus.rst.rrt",
    "spa.rst.rststb",
    "spa.rst.sctb",
    "tha.pdtb.tdtb",
    "tur.pdtb.tdb",
    "tur.pdtb.tedm",
    "zho.dep.scidtb",
    "zho.pdtb.cdtb",
    "zho.rst.gcdt",
    "zho.rst.sctb"
]
def fix_mwe(sentence):
    mwe={}
    sentence['parent_mwe']=[]
    for i, x in enumerate(sentence['id']):
        if '-' in x:
            for a in x.split('-'):
                mwe[a]=sentence['form'][i]
        sentence['parent_mwe']+=[mwe.get(x,'')]

    for i, x in enumerate(sentence['id']):
        if "-" in x:
            for k,v in sentence.items():
                del v[i]
    return sentence

def parse_conll_stream(file_stream):
    names = ['id', 'form', 'lemma', 'upos', 'xpos', 'feats', 'head', 'deprel', 'deps', 'misc','doc_id']
    sentence = {name: [] for name in names}
    mwe_id=[]
    for line in file_stream:
        line = line.strip()
        if line.startswith("#"):
            if "doc_id" in line:
                doc_id=line.split('=')[-1].strip()
            continue
        if not line:
            if sentence['id']:
                yield sentence
                sentence = {name: [] for name in names}
            continue
        token_data = line.split('\t') + [doc_id]
        for name, value in zip(names, token_data):
            if name=='id' and not value.isnumeric():
                mwe_id=value.split('-')                
            else:
                sentence[name].append(value)

def get_kwarg_names(func):
    return [k for k, v in inspect.signature(func).parameters.items() if v.default != v.empty]
    
_URLs = {f'{task}-{split}.{type}':f"https://raw.githubusercontent.com/disrpt/sharedtask2023/main/data/{task}/{task}_{split}.{type}" \
         for task in files for split in 'train dev test'.split() for type in ['rels','conllu']}
#_URLs = {k:v for k,v in _URLs.items() if requests.get(v).status_code!=404}

conllu_features = ['id', 'form', 'lemma', 'upos', 'xpos', 'feats', 'head', 'deprel', 'deps', 'misc', 'seg','doc_id']
feature_type = {"seg":datasets.features.Sequence(
                        datasets.features.ClassLabel(names=["O","B-Segment"])),
                'id':datasets.Value("string"),'doc_id':datasets.Value("string")}

conllu_features = datasets.Features({x:feature_type.get(x,datasets.Sequence(datasets.Value("string")))
                                     for x in conllu_features})
 
def map_seg(x):
    return [("B-Segment" if "beginseg=yes" in a.lower() else "O") for a in x]

def remove_type(x):
    return x.replace(".rels","").replace(".conllu","")

class Dataset(datasets.GeneratorBasedBuilder):
    
    BUILDER_CONFIGS = [
            Config(
                name=f"{n}.{type}",
                data_dir=f"{n}.{type}",
            ) for n in files for type in ["rels","conllu"]
    ]
    def __init__(self,*args,**kwargs):
        self.BUILDER_CONFIG_CLASS.__post_init__=lambda x:x
        base_kwargs_names=get_kwarg_names(super().__init__)
        gen_kwargs={}
        self.files={}
        self.preprocessed_underscores=dict()
        for k,v in copy.deepcopy(kwargs).items():
            if k not in base_kwargs_names:
                gen_kwargs[k]=v
                del kwargs[k]
        self.gen_kwargs=gen_kwargs
        return super().__init__(*args,**kwargs)
        
    def _split_generators(self, dl_manager: datasets.DownloadManager):
        cfg_name = self.config.name.rsplit('.', 1)[0]
        data_dir = remove_type(self.config.data_dir)
        type = self.config.name.split('.')[-1]
        urls={k:v for (k,v) in _URLs.items() if cfg_name in k and requests.get(v).status_code!=404}
        data_file = dl_manager.download(urls)
        self.files = {**self.files, **data_file}

        splits_dict = {datasets.Split.TRAIN: 'train', datasets.Split.VALIDATION: 'dev', datasets.Split.TEST: 'test'}

        split_generators = [
            datasets.SplitGenerator(name=split, gen_kwargs={"filepath": data_file[f"{data_dir}-{key}.{type}"]})
            for split, key in splits_dict.items()
            if f"{data_dir}-{key}.{type}" in data_file
        ]
        return split_generators

    def _info(self): return datasets.DatasetInfo(
        citation=key_to_entry.get(datasets_and_citations.get(remove_type(self.config.name)),None),
        features=(None if ".rels" in self.config.name else conllu_features)
    )
        
    def _generate_examples(self, filepath):
        print(filepath)
        corpus=self.config.name.split('.')[2]
        run_args={
        'corpus':corpus,
        'rel_files': [v for k, v in self.files.items() if '.rels' in k],
        'dep_files': [v for k, v in self.files.items() if '.conllu' in k],
        **{k:v for k,v in self.gen_kwargs.items() if 'path' in k}
        }
        print('run_args',run_args)
        if corpus in ['rstdt','pdtb','cdtb','gum','tdb'] and not self.preprocessed_underscores.get(corpus,False) and self.gen_kwargs.get('process_underscore',True):
            run(**run_args)
            self.preprocessed_underscores[corpus]=True
            
        with open(filepath, encoding="utf-8") as f:
            if "conllu" in self.config.name:
                stream=parse_conll_stream(f)
                for i, row in enumerate(stream):
                    row['seg']=map_seg(row['misc'])
                    row['doc_id']=row['doc_id'][0]
                    yield i,row
            reader = csv.DictReader(f,delimiter='\t',quoting=csv.QUOTE_NONE)
            for id_, row in enumerate(reader):
                if id_ == 0:
                    continue
                yield id_, row