David Wadden commited on
Commit
fd02f3e
1 Parent(s): 9c66dd8

Fix the entailment script.

Browse files
Files changed (1) hide show
  1. scifact_entailment.py +58 -47
scifact_entailment.py CHANGED
@@ -2,9 +2,8 @@
2
  using evidence from the cited abstracts. Formatted as a paragraph-level entailment task."""
3
 
4
 
5
- import json
6
-
7
  import datasets
 
8
 
9
 
10
  _CITATION = """\
@@ -20,6 +19,12 @@ _DESCRIPTION = """\
20
  SciFact, a dataset of 1.4K expert-written scientific claims paired with evidence-containing abstracts, and annotated with labels and rationales.
21
  """
22
 
 
 
 
 
 
 
23
 
24
  class ScifactEntailmentConfig(datasets.BuilderConfig):
25
  """BuilderConfig for Scifact"""
@@ -43,17 +48,15 @@ class ScifactEntailment(datasets.GeneratorBasedBuilder):
43
 
44
  def _info(self):
45
  # TODO(scifact): Specifies the datasets.DatasetInfo object
 
46
  features = {
47
- "id": datasets.Value("int32"), # An integer claim ID.
48
- "claim": datasets.Value("string"), # The text of the claim.
49
- "evidence_doc_id": datasets.Value("string"),
50
- "evidence_label": datasets.Value("string"), # Label for the rationale.
51
- "evidence_sentences": datasets.features.Sequence(
52
- datasets.Value("int32")
53
- ), # Rationale sentences.
54
- "cited_doc_ids": datasets.features.Sequence(
55
- datasets.Value("int32")
56
- ), # The claim's "cited documents".
57
  }
58
 
59
  return datasets.DatasetInfo(
@@ -73,74 +76,82 @@ class ScifactEntailment(datasets.GeneratorBasedBuilder):
73
  citation=_CITATION,
74
  )
75
 
 
 
 
 
 
 
 
 
 
76
  def _split_generators(self, dl_manager):
77
  """Returns SplitGenerators."""
78
  # TODO(scifact): Downloads the data and defines the splits
79
  # dl_manager is a datasets.download.DownloadManager that can be used to
80
  # download and extract URLs
 
 
 
 
 
 
 
 
 
 
81
  return [
82
  datasets.SplitGenerator(
83
  name=datasets.Split.TRAIN,
84
  # These kwargs will be passed to _generate_examples
85
  gen_kwargs={
 
 
86
  "split": "train",
87
  },
88
  ),
89
- datasets.SplitGenerator(
90
- name=datasets.Split.TEST,
91
- # These kwargs will be passed to _generate_examples
92
- gen_kwargs={
93
- "split": "test",
94
- },
95
- ),
96
  datasets.SplitGenerator(
97
  name=datasets.Split.VALIDATION,
98
  # These kwargs will be passed to _generate_examples
99
  gen_kwargs={
100
- "split": "dev",
 
 
101
  },
102
  ),
103
  ]
104
 
105
- def _generate_examples(self, split):
106
  """Yields examples."""
107
- # TODO(scifact): Yields (key, example) tuples from the dataset
108
-
109
- # Load corpus and convert to dict.
110
- corpus = datasets.load_dataset("bigbio/scifact", "scifact_corpus_source", split="train")
111
- corpus = {x["doc_id"]: x for x in corpus}
112
-
113
- # Load claims.
114
- claims = datasets.load_dataset("bigbio/scifact", "scifact_claims_source", split=split)
115
-
116
- for id_, claim in enumerate(claims):
117
- evidence = {x["doc_id"]: x for x in claim["evidences"]}
118
  for cited_doc_id in claim["cited_doc_ids"]:
119
  cited_doc = corpus[cited_doc_id]
120
- # Format the abstract.
121
- sent_ids = [f"[{i}]" for i in range(len(cited_doc["abstract"]))]
122
- # Get rid of newlines.
123
- sents = [sent.strip() for sent in cited_doc["abstract"]]
124
- zipped = zip(sent_ids, sents)
125
- cited_abstract = " ".join(
126
- [f"{entry[0]} {entry[1]}" for entry in zipped]
127
- )
128
 
129
  if cited_doc_id in evidence:
130
- verdict = evidence[cited_doc_id]["label"]
131
- sents = evidence[cited_doc_id]["sentence_ids"]
 
 
 
 
 
132
  else:
133
  verdict = "NEI"
134
- sents = []
135
 
136
  instance = {
137
- "id": claim["id"],
138
  "claim": claim["claim"],
139
  "abstract_id": cited_doc_id,
140
- "title": cited_doc["title"],
141
- "abstract": cited_abstract,
142
  "verdict": verdict,
143
- "evidence": sents,
144
  }
145
 
 
146
  yield id_, instance
 
2
  using evidence from the cited abstracts. Formatted as a paragraph-level entailment task."""
3
 
4
 
 
 
5
  import datasets
6
+ import json
7
 
8
 
9
  _CITATION = """\
 
19
  SciFact, a dataset of 1.4K expert-written scientific claims paired with evidence-containing abstracts, and annotated with labels and rationales.
20
  """
21
 
22
+ _URL = "https://scifact.s3-us-west-2.amazonaws.com/release/latest/data.tar.gz"
23
+
24
+
25
+ def flatten(xss):
26
+ return [x for xs in xss for x in xs]
27
+
28
 
29
  class ScifactEntailmentConfig(datasets.BuilderConfig):
30
  """BuilderConfig for Scifact"""
 
48
 
49
  def _info(self):
50
  # TODO(scifact): Specifies the datasets.DatasetInfo object
51
+
52
  features = {
53
+ "claim_id": datasets.Value("int32"),
54
+ "claim": datasets.Value("string"),
55
+ "abstract_id": datasets.Value("int32"),
56
+ "title": datasets.Value("string"),
57
+ "abstract": datasets.features.Sequence(datasets.Value("string")),
58
+ "verdict": datasets.Value("string"),
59
+ "evidence": datasets.features.Sequence(datasets.Value("int32")),
 
 
 
60
  }
61
 
62
  return datasets.DatasetInfo(
 
76
  citation=_CITATION,
77
  )
78
 
79
+ @staticmethod
80
+ def _read_tar_file(f):
81
+ res = []
82
+ for row in f:
83
+ this_row = json.loads(row.decode("utf-8"))
84
+ res.append(this_row)
85
+
86
+ return res
87
+
88
  def _split_generators(self, dl_manager):
89
  """Returns SplitGenerators."""
90
  # TODO(scifact): Downloads the data and defines the splits
91
  # dl_manager is a datasets.download.DownloadManager that can be used to
92
  # download and extract URLs
93
+ archive = dl_manager.download(_URL)
94
+ for path, f in dl_manager.iter_archive(archive):
95
+ if path == "data/corpus.jsonl":
96
+ corpus = self._read_tar_file(f)
97
+ corpus = {x["doc_id"]: x for x in corpus}
98
+ elif path == "data/claims_train.jsonl":
99
+ claims_train = self._read_tar_file(f)
100
+ elif path == "data/claims_dev.jsonl":
101
+ claims_validation = self._read_tar_file(f)
102
+
103
  return [
104
  datasets.SplitGenerator(
105
  name=datasets.Split.TRAIN,
106
  # These kwargs will be passed to _generate_examples
107
  gen_kwargs={
108
+ "claims": claims_train,
109
+ "corpus": corpus,
110
  "split": "train",
111
  },
112
  ),
 
 
 
 
 
 
 
113
  datasets.SplitGenerator(
114
  name=datasets.Split.VALIDATION,
115
  # These kwargs will be passed to _generate_examples
116
  gen_kwargs={
117
+ "claims": claims_validation,
118
+ "corpus": corpus,
119
+ "split": "validation",
120
  },
121
  ),
122
  ]
123
 
124
+ def _generate_examples(self, claims, corpus, split):
125
  """Yields examples."""
126
+ # Loop over claims and put evidence together with claim.
127
+ id_ = -1 # Will increment to 0 on first iteration.
128
+ for claim in claims:
129
+ evidence = {int(k): v for k, v in claim["evidence"].items()}
 
 
 
 
 
 
 
130
  for cited_doc_id in claim["cited_doc_ids"]:
131
  cited_doc = corpus[cited_doc_id]
132
+ abstract_sents = [sent.strip() for sent in cited_doc["abstract"]]
 
 
 
 
 
 
 
133
 
134
  if cited_doc_id in evidence:
135
+ this_evidence = evidence[cited_doc_id]
136
+ verdict = this_evidence[0][
137
+ "label"
138
+ ] # Can take first evidence since all labels are same.
139
+ evidence_sents = flatten(
140
+ [entry["sentences"] for entry in this_evidence]
141
+ )
142
  else:
143
  verdict = "NEI"
144
+ evidence_sents = []
145
 
146
  instance = {
147
+ "claim_id": claim["id"],
148
  "claim": claim["claim"],
149
  "abstract_id": cited_doc_id,
150
+ "title": cited_doc["title"],
151
+ "abstract": abstract_sents,
152
  "verdict": verdict,
153
+ "evidence": evidence_sents,
154
  }
155
 
156
+ id_ += 1
157
  yield id_, instance