vblagoje commited on
Commit
19f6b93
1 Parent(s): e2291eb

Add passage generation

Browse files

Files changed (1) hide show
  1. wikipedia_snippets_streamed.py +51 -7
wikipedia_snippets_streamed.py CHANGED
@@ -1,5 +1,5 @@
1
  """Wikipedia snippets in parquet format"""
2
- import os
3
  import pyarrow.parquet as pq
4
  import datasets
5
 
@@ -25,16 +25,20 @@ _LICENSE = (
25
 
26
 
27
  class WikipediaSnippetsStreamed(datasets.GeneratorBasedBuilder):
28
- """Bengali wikipedia from 03/20/2021"""
29
 
30
  def _info(self):
31
  return datasets.DatasetInfo(
32
  description=_DESCRIPTION,
33
  features=datasets.Features(
34
  {
35
- "wikidata_id": datasets.Value("string"),
36
- "text": datasets.Value("string"),
37
- "version_id": datasets.Value("string"),
 
 
 
 
 
38
  }
39
  ),
40
  supervised_keys=None,
@@ -49,10 +53,50 @@ class WikipediaSnippetsStreamed(datasets.GeneratorBasedBuilder):
49
  datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": downloaded_file}),
50
  ]
51
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
52
  def _generate_examples(self, filepath):
53
  logger.info("generating examples from = %s", filepath)
 
54
  with open(filepath, "rb") as f:
55
  pf = pq.ParquetFile(f)
56
  for i in range(pf.num_row_groups):
57
- id_ = f"{filepath}_{i}"
58
- yield id_, pf.read_row_group(i).to_pydict()
 
 
 
 
1
  """Wikipedia snippets in parquet format"""
2
+ import math
3
  import pyarrow.parquet as pq
4
  import datasets
5
 
25
 
26
 
27
  class WikipediaSnippetsStreamed(datasets.GeneratorBasedBuilder):
 
28
 
29
  def _info(self):
30
  return datasets.DatasetInfo(
31
  description=_DESCRIPTION,
32
  features=datasets.Features(
33
  {
34
+ "wiki_id": datasets.Value("string"),
35
+ "start_paragraph": datasets.Value("int32"),
36
+ "start_character": datasets.Value("int32"),
37
+ "end_paragraph": datasets.Value("int32"),
38
+ "end_character": datasets.Value("int32"),
39
+ "article_title": datasets.Value("string"),
40
+ "section_title": datasets.Value("string"),
41
+ "passage_text": datasets.Value("string"),
42
  }
43
  ),
44
  supervised_keys=None,
53
  datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": downloaded_file}),
54
  ]
55
 
56
+ def wiki40b_article_snippets(self, article, passage_len=100, overlap=0):
57
+ paragraphs = article["text"].split("\n")
58
+ aticle_idx = paragraphs.index("_START_ARTICLE_") + 1
59
+ article_title = paragraphs[aticle_idx] if aticle_idx < len(paragraphs) else ""
60
+ section_indices = [i + 1 for i, par in enumerate(paragraphs[:-1]) if par == "_START_SECTION_"]
61
+ par_tabs = [par.split(" ") for par in paragraphs]
62
+ word_map = [
63
+ (i, len(" ".join(par[:j])), w)
64
+ for i, par in enumerate(par_tabs)
65
+ if not par[0].startswith("_START_")
66
+ for j, w in enumerate(par)
67
+ if i > 0
68
+ ]
69
+ step_size = passage_len - overlap
70
+ passages = []
71
+ for i in range(math.ceil(len(word_map) / step_size)):
72
+ pre_toks = word_map[i * step_size: i * step_size + passage_len]
73
+ start_section_id = max([0] + [j for j in section_indices if j <= pre_toks[0][0]])
74
+ section_ids = [j for j in section_indices if j >= start_section_id and j <= pre_toks[-1][0]]
75
+ section_ids = section_ids if len(section_ids) > 0 else [0]
76
+ passage_text = " ".join([w for p_id, s_id, w in pre_toks])
77
+ passages += [
78
+ {
79
+ "article_title": article_title,
80
+ "section_title": " & ".join([paragraphs[j] for j in section_ids]),
81
+ "wiki_id": article["wikidata_id"],
82
+ "start_paragraph": pre_toks[0][0],
83
+ "start_character": pre_toks[0][1],
84
+ "end_paragraph": pre_toks[-1][0],
85
+ "end_character": pre_toks[-1][1] + len(pre_toks[-1][2]) + 1,
86
+ "passage_text": passage_text.replace("_NEWLINE_", "\n"),
87
+ }
88
+ ]
89
+ return passages
90
+
91
  def _generate_examples(self, filepath):
92
  logger.info("generating examples from = %s", filepath)
93
+ passage_counter = 0
94
  with open(filepath, "rb") as f:
95
  pf = pq.ParquetFile(f)
96
  for i in range(pf.num_row_groups):
97
+ batch_table_dict = pf.read_row_group(i).to_pydict()
98
+ for idx, article_text in enumerate(batch_table_dict["text"]):
99
+ article = {"text": article_text, "wikidata_id": batch_table_dict["wikidata_id"][idx]}
100
+ passages = self.wiki40b_article_snippets(article)
101
+ for passage in passages:
102
+ yield ++passage_counter, passage