File size: 5,332 Bytes
35754f0 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 | """Small original-source conversions that avoid audio downloads or legacy loaders."""
from collections import Counter
import hashlib
import io
import json
from pathlib import Path
import urllib.request
import xml.etree.ElementTree as ET
import zipfile
from datasets import Dataset, DatasetDict
from .relevance import text_key, write_mirror
SLURP_REVISION = '8eb16545762be97ace75334109d73824217311f1'
MCSCRIPT_URL = 'https://fedora.clarin-d.uni-saarland.de/sfb1102/MCScript-2.0.zip'
MCSCRIPT_SHA256 = '50037b08123a4660e2e00ad07bcfa2a7ddc1d666b05b198d359cd2cda8f5e88c'
def slurp():
"""SLURP original human transcripts and intent labels, without bundled audio.
Native real train/devel/test partitions; synthetic speech is not included.
Identical transcripts with inconsistent intents are excluded, and duplicate
transcripts stay in their highest-priority native holdout (test, dev, train).
One- and two-word transcripts are excluded as insufficient intent context.
Original terms: CC BY 4.0. All removals and source hashes are recorded.
"""
splits={}; hashes={}; labels={};counts=Counter()
for split,native in [('train','train'),('validation','devel'),('test','test')]:
url=f'https://raw.githubusercontent.com/pswietojanski/slurp/{SLURP_REVISION}/dataset/slurp/{native}.jsonl'
raw=urllib.request.urlopen(url).read();hashes[native]=hashlib.sha256(raw).hexdigest()
splits[split]=[json.loads(line) for line in raw.splitlines() if line.strip()]
for row in splits[split]:labels.setdefault(text_key(row['sentence']),set()).add(row['intent'])
output={};seen=set()
for split in ('test','validation','train'):
output[split]=[]
for row in splits[split]:
key=text_key(row['sentence'])
reason='ambiguous_intent' if len(labels[key])!=1 else 'short_transcript' if len(row['sentence'].split())<3 else 'duplicate_transcript' if key in seen else None
if reason:counts[split+'/'+reason]+=1;continue
seen.add(key)
meta=dict(id=str(row['slurp_id']),split_group_id=key,scenario=row['scenario'],entities=row['entities'],source_revision=SLURP_REVISION)
output[split].append(dict(sentence1=row['sentence'],sentence2='',labels=row['intent'],metadata=json.dumps(meta,sort_keys=True)))
return write_mirror('slurp',output,'string',dict(source='https://github.com/pswietojanski/slurp',revision=SLURP_REVISION,
source_hashes=hashes,excluded=dict(counts),split_policy='native splits; exact transcript conflicts and overlap removed',
conversion_sha256=hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),license='cc-by-4.0'))
def mcscript2():
"""MCScript2.0 native two-answer reading comprehension, original XML partitions.
Preserve stories, questions and native choices; exactly one gold answer and
two distinct non-empty options are required. Story hashes prevent train/dev/
test copies. Source URL/archive hash and per-split counts are recorded.
Dataset terms: unspecified until original distribution terms are verified.
"""
raw=urllib.request.urlopen(MCSCRIPT_URL).read()
if hashlib.sha256(raw).hexdigest()!=MCSCRIPT_SHA256:raise ValueError('MCScript archive changed')
output={};groups={};counts=Counter()
with zipfile.ZipFile(io.BytesIO(raw)) as archive:
for split,native in [('train','train'),('validation','dev'),('test','test')]:
output[split]=[];groups[split]=set()
for story in ET.fromstring(archive.read(native+'-data.xml')):
text=story.findtext('text');group=text_key(text);groups[split].add(group)
for question in story.findall('./questions/question'):
choices=[a.attrib['text'] for a in question.findall('answer')]
gold=[i for i,a in enumerate(question.findall('answer')) if a.attrib['correct'].lower()=='true']
if len(choices)!=2 or len(set(choices))!=2 or not all(c.strip() for c in choices) or len(gold)!=1:
counts[split+'/invalid_choices']+=1;continue
meta=dict(id=story.attrib['id']+':'+question.attrib['id'],split_group_id=group,
scenario=story.attrib['scenario'],question_type=question.attrib['type'])
output[split].append(dict(inputs=f"Story: {text}\nQuestion: {question.attrib['text']}",
choices_list=choices,labels=gold[0],metadata=json.dumps(meta,sort_keys=True)))
assert all(not groups[a]&groups[b] for a in groups for b in groups if a!=b),'MCScript story overlap'
data=DatasetDict({s:Dataset.from_list(r) for s,r in output.items()})
directory=Path('build/mcscript2-release');directory.mkdir(parents=True,exist_ok=True)
for split,rows in data.items():rows.to_parquet(str(directory/(split+'.parquet')))
provenance=dict(source_url=MCSCRIPT_URL,archive_sha256=MCSCRIPT_SHA256,excluded=dict(counts),
splits={s:len(r) for s,r in data.items()},split_policy='native story partitions',
conversion_sha256=hashlib.sha256(Path(__file__).read_bytes()).hexdigest(),license='unspecified')
(directory/'provenance.json').write_text(json.dumps(provenance,indent=2)+'\n')
(directory/'excluded-questions.jsonl').write_text('')
return data
|