from https://github.com/ArneBinder/pie-datasets/pull/103

#3
Files changed (5) hide show
  1. README.md +69 -38
  2. img/leaannof3.png +3 -0
  3. img/sciarg-sam.png +3 -0
  4. requirements.txt +2 -1
  5. sciarg.py +100 -24
README.md CHANGED
@@ -22,13 +22,18 @@ The language in the dataset is English (scientific academic publications on comp
22
 
23
  ### Dataset Variants
24
 
25
- The `sciarg` dataset comes in a single version (`default`) with `BratDocumentWithMergedSpans` as document type. Note,
26
- that this in contrast to the base `brat` dataset, where the document type for the `default` variant is `BratDocument`.
27
- The reason is that the SciArg dataset was published with spans that are just fragmented by whitespace which seems
 
 
28
  to be because of the annotation tool used. In the `sciarg` dataset, we merge these fragments, so that the document type
29
- can be `BratDocumentWithMergedSpans` (this is easier to handle for most of the task modules). However, fragmented
30
- spans are conceptually also available in SciArg, but they are marked with the `parts_of_same` relation which are kept
31
- as they are in the `sciarg` (`default`) dataset.
 
 
 
32
 
33
  ### Data Schema
34
 
@@ -42,41 +47,28 @@ from pie_datasets import load_dataset, builders
42
  # load default version
43
  datasets = load_dataset("pie/sciarg")
44
  doc = datasets["train"][0]
45
- assert isinstance(doc, builders.brat.BratDocument)
46
 
47
- # load version with merged span fragments
48
- dataset_merged_spans = load_dataset("pie/sciarg", name="merge_fragmented_spans")
49
- doc_merged_spans = dataset_merged_spans["train"][0]
50
- assert isinstance(doc_merged_spans, builders.brat.BratDocumentWithMergedSpans)
51
  ```
52
 
53
- ### Document Converters
54
-
55
- The dataset provides document converters for the following target document types:
56
-
57
- - `pytorch_ie.documents.TextDocumentWithLabeledSpansAndBinaryRelations`
58
- - `LabeledSpans`, converted from `BratDocument`'s `spans`
59
- - labels: `background_claim`, `own_claim`, `data`
60
- - if `spans` contain whitespace at the beginning and/or the end, the whitespace are trimmed out.
61
- - `BinraryRelations`, converted from `BratDocument`'s `relations`
62
- - labels: `supports`, `contradicts`, `semantically_same`, `parts_of_same`
63
- - if the `relations` label is `semantically_same` or `parts_of_same`, they are merged if they are the same arguments after sorting.
64
- - `pytorch_ie.documents.TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions`
65
- - `LabeledSpans`, as above
66
- - `BinaryRelations`, as above
67
- - `LabeledPartitions`, partitioned `BratDocument`'s `text`, according to the paragraph, using regex.
68
- - labels: `title`, `abstract`, `H1`
69
-
70
- See [here](https://github.com/ChristophAlt/pytorch-ie/blob/main/src/pytorch_ie/documents.py) for the document type
71
- definitions.
72
-
73
  ### Data Splits
74
 
75
  The dataset consists of a single `train` split that has 40 documents.
76
 
77
  For detailed statistics on the corpus, see Lauscher et al. ([2018](<(https://aclanthology.org/W18-5206/)>), p. 43), and the author's [resource analysis](https://github.com/anlausch/sciarg_resource_analysis).
78
 
79
- ### Label Descriptions
 
 
 
 
 
 
 
80
 
81
  #### Components
82
 
@@ -113,18 +105,57 @@ For detailed statistics on the corpus, see Lauscher et al. ([2018](<(https://acl
113
 
114
  - `semantically_same`: between two mentions of effectively the same claim or data component. Can be seen as *argument coreference*, analogous to entity, and *event coreference*. This relation is considered symmetric (i.e., **bidirectional**) and non-argumentative.
115
  (Lauscher et al. 2018, p.41; following [Dung, 1995](https://www.sciencedirect.com/science/article/pii/000437029400041X?via%3Dihub))
116
- - `parts_of_same`: when a single component is split up in several parts. It is **non-argumentative**, **bidirectional**, but also **intra-component**
117
 
118
  (*Annotation Guidelines*, pp. 4-6)
119
 
120
- **Important note on label counts**:
121
 
122
- There are currently discrepancies in label counts between
123
 
124
- - previous report in [Lauscher et al., 2018](https://aclanthology.org/W18-5206/), p. 43),
125
- - current report above here (labels counted in `BratDocument`'s);
126
 
127
- possibly since [Lauscher et al., 2018](https://aclanthology.org/W18-5206/) presents the numbers of the real argumentative components, whereas here discontinuous components are still split (marked with the `parts_of_same` helper relation) and, thus, count per fragment.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
128
 
129
  ## Dataset Creation
130
 
 
22
 
23
  ### Dataset Variants
24
 
25
+ The `sciarg` dataset comes in two versions: `default` and `resolve_parts_of_same`.
26
+
27
+ First, the `default` version with `BratDocumentWithMergedSpans` as document type.
28
+ In contrast to the base `brat` dataset, where the document type for the `default` variant is `BratDocument`,
29
+ the SciArg dataset was published with spans that are just fragmented by whitespace which seems
30
  to be because of the annotation tool used. In the `sciarg` dataset, we merge these fragments, so that the document type
31
+ can be `BratDocumentWithMergedSpans` (this is easier to handle for most of the task modules).
32
+ Fragmented spans, which belong to the same argumentative unit, are marked with `parts_of_same` relations.
33
+
34
+ Second, the `resolve_parts_of_same` version with `BratDocument` as document type.
35
+ In this version, all fragmented spans which were separated by other argumentative or non-argumentative spans and
36
+ are connected via the `parts_of_same` relations are converted to `LabeledMultiSpans`.
37
 
38
  ### Data Schema
39
 
 
47
  # load default version
48
  datasets = load_dataset("pie/sciarg")
49
  doc = datasets["train"][0]
50
+ assert isinstance(doc, builders.brat.BratDocumentWithMergedSpans)
51
 
52
+ # load version with resolved parts_of_same relations
53
+ datasets = load_dataset("pie/sciarg", name='resolve_parts_of_same')
54
+ doc = datasets["train"][0]
55
+ assert isinstance(doc, builders.brat.BratDocument)
56
  ```
57
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58
  ### Data Splits
59
 
60
  The dataset consists of a single `train` split that has 40 documents.
61
 
62
  For detailed statistics on the corpus, see Lauscher et al. ([2018](<(https://aclanthology.org/W18-5206/)>), p. 43), and the author's [resource analysis](https://github.com/anlausch/sciarg_resource_analysis).
63
 
64
+ ### Label Descriptions and Statistics
65
+
66
+ In this section, we report our own corpus' statistics; however, there are currently discrepancies in label counts between our report and:
67
+
68
+ - previous report in [Lauscher et al., 2018](https://aclanthology.org/W18-5206/), p. 43),
69
+ - current report above here (labels counted in `BratDocumentWithMergedSpans`'s);
70
+
71
+ possibly since [Lauscher et al., 2018](https://aclanthology.org/W18-5206/) presents the numbers of the real argumentative components, whereas here discontinuous components are still split (marked with the `parts_of_same` helper relation) and, thus, count per fragment.
72
 
73
  #### Components
74
 
 
105
 
106
  - `semantically_same`: between two mentions of effectively the same claim or data component. Can be seen as *argument coreference*, analogous to entity, and *event coreference*. This relation is considered symmetric (i.e., **bidirectional**) and non-argumentative.
107
  (Lauscher et al. 2018, p.41; following [Dung, 1995](https://www.sciencedirect.com/science/article/pii/000437029400041X?via%3Dihub))
108
+ - `parts_of_same` (only in the `default` dataset variant): when a single component is split up in several parts. It is **non-argumentative**, **bidirectional**, but also **intra-component**
109
 
110
  (*Annotation Guidelines*, pp. 4-6)
111
 
112
+ #### Examples
113
 
114
+ ![sample1](img/leaannof3.png)
115
 
116
+ Above: Diagram from *Annotation Guildelines* (p.6)
 
117
 
118
+ Below: Subset of relations in `A01`
119
+
120
+ ![sample2](img/sciarg-sam.png)
121
+
122
+ ### Document Converters
123
+
124
+ The dataset provides document converters for the following target document types:
125
+
126
+ From `default` version:
127
+
128
+ - `pie_modules.documents.TextDocumentWithLabeledSpansAndBinaryRelations`
129
+ - `labeled_spans`: `LabeledSpan` annotations, converted from `BratDocumentWithMergedSpans`'s `spans`
130
+ - labels: `background_claim`, `own_claim`, `data`
131
+ - if `spans` contain whitespace at the beginning and/or the end, that whitespace is trimmed out.
132
+ - `binary_relations`: `BinaryRelation` annotations, converted from `BratDocumentWithMergedSpans`'s `relations`
133
+ - labels: `supports`, `contradicts`, `semantically_same`, `parts_of_same`
134
+ - if the `relations` label is `semantically_same` or `parts_of_same` (i.e. it is a symmetric relation), their arguments are sorted by their start and end indices.
135
+ - `pie_modules.documents.TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions`
136
+ - `labeled_spans`, as above
137
+ - `binary_relations`, as above
138
+ - `labeled_partitions`, `LabeledSpan` annotations, created from splitting `BratDocumentWithMergedSpans`'s `text` at new paragraph in `xml` format.
139
+ - labels: `title`, `abstract`, `H1`
140
+
141
+ From `resolve_parts_of_same` version:
142
+
143
+ - `pie_modules.documents.TextDocumentWithLabeledMultiSpansAndBinaryRelations`:
144
+ - `labeled_multi_spans`: `LabeledMultiSpan` annotations, converted from `BratDocument`'s `spans`
145
+ - labels: as above
146
+ - if spans contain whitespace at the beginning and/or the end, that whitespace is trimmed out.
147
+ - `binary_relations`: `BinaryRelation` annotations, converted from `BratDocument`'s `relations`
148
+ - labels: `supports`, `contradicts`, `semantically_same`
149
+ - in contrast to the `default` version, spans connected with `parts_of_same` relation are stored as one labeled multi-span
150
+ - if the `relations` label is `semantically_same` (i.e. it is a symmetric relation), their arguments are sorted by their start and end indices.
151
+ - `pie_modules.documents.TextDocumentWithLabeledMultiSpansBinaryRelationsAndLabeledPartitions`:
152
+ - `labeled_multi_spans`, as above
153
+ - `binary_relations`, as above
154
+ - `labeled_partitions`, `LabeledSpan` annotations, created from splitting `BratDocument`'s `text` at new paragraph in `xml` format.
155
+ - labels: `title`, `abstract`, `H1`
156
+
157
+ See [here](https://github.com/ArneBinder/pie-modules/blob/main/src/pie_modules/documents.py) for the document type
158
+ definitions.
159
 
160
  ## Dataset Creation
161
 
img/leaannof3.png ADDED

Git LFS Details

  • SHA256: fc5934d1640016e6bf87ecde76da735c0d1a6deab6e39ae89032b2461d9d7ab6
  • Pointer size: 130 Bytes
  • Size of remote file: 38 kB
img/sciarg-sam.png ADDED

Git LFS Details

  • SHA256: a1104907623460a5517b4968c00386cd71c736dcca7c2fd551b0d2f8b6dc2cea
  • Pointer size: 130 Bytes
  • Size of remote file: 49.1 kB
requirements.txt CHANGED
@@ -1,2 +1,3 @@
1
  pie-datasets>=0.6.0,<0.9.0
2
- pie-modules>=0.8.0,<0.9.0
 
 
1
  pie-datasets>=0.6.0,<0.9.0
2
+ pie-modules>=0.10.8,<0.11.0
3
+ networkx>=3.0.0,<4.0.0
sciarg.py CHANGED
@@ -1,23 +1,27 @@
1
  from pie_modules.document.processing import (
2
  RegexPartitioner,
3
  RelationArgumentSorter,
 
4
  TextSpanTrimmer,
5
  )
6
- from pytorch_ie.core import Document
7
- from pytorch_ie.documents import (
 
8
  TextDocumentWithLabeledSpansAndBinaryRelations,
9
  TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions,
10
  )
 
11
 
12
  from pie_datasets.builders import BratBuilder, BratConfig
13
- from pie_datasets.builders.brat import BratDocumentWithMergedSpans
 
14
  from pie_datasets.document.processing import Caster, Pipeline
15
 
16
  URL = "http://data.dws.informatik.uni-mannheim.de/sci-arg/compiled_corpus.zip"
17
  SPLIT_PATHS = {"train": "compiled_corpus"}
18
 
19
 
20
- def get_common_pipeline_steps(target_document_type: type[Document]) -> dict:
21
  return dict(
22
  cast=Caster(
23
  document_type=target_document_type,
@@ -31,6 +35,33 @@ def get_common_pipeline_steps(target_document_type: type[Document]) -> dict:
31
  )
32
 
33
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
34
  class SciArg(BratBuilder):
35
  BASE_DATASET_PATH = "DFKI-SLT/brat"
36
  BASE_DATASET_REVISION = "844de61e8a00dc6a93fc29dc185f6e617131fbf1"
@@ -39,33 +70,78 @@ class SciArg(BratBuilder):
39
  # The span fragments in SciArg come just from the new line splits, so we can merge them.
40
  # Actual span fragments are annotated via "parts_of_same" relations.
41
  BUILDER_CONFIGS = [
42
- BratConfig(name=BratBuilder.DEFAULT_CONFIG_NAME, merge_fragmented_spans=True),
 
43
  ]
44
  DOCUMENT_TYPES = {
45
  BratBuilder.DEFAULT_CONFIG_NAME: BratDocumentWithMergedSpans,
 
46
  }
47
 
48
  # we need to add None to the list of dataset variants to support the default dataset variant
49
  BASE_BUILDER_KWARGS_DICT = {
50
  dataset_variant: {"url": URL, "split_paths": SPLIT_PATHS}
51
- for dataset_variant in ["default", "merge_fragmented_spans", None]
52
  }
53
 
54
- DOCUMENT_CONVERTERS = {
55
- TextDocumentWithLabeledSpansAndBinaryRelations: Pipeline(
56
- **get_common_pipeline_steps(TextDocumentWithLabeledSpansAndBinaryRelations)
57
- ),
58
- TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions: Pipeline(
59
- **get_common_pipeline_steps(
60
- TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions
61
- ),
62
- add_partitions=RegexPartitioner(
63
- partition_layer_name="labeled_partitions",
64
- pattern="<([^>/]+)>.*</\\1>",
65
- label_group_id=1,
66
- label_whitelist=["Title", "Abstract", "H1"],
67
- skip_initial_partition=True,
68
- strip_whitespace=True,
69
- ),
70
- ),
71
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  from pie_modules.document.processing import (
2
  RegexPartitioner,
3
  RelationArgumentSorter,
4
+ SpansViaRelationMerger,
5
  TextSpanTrimmer,
6
  )
7
+ from pie_modules.documents import (
8
+ TextDocumentWithLabeledMultiSpansAndBinaryRelations,
9
+ TextDocumentWithLabeledMultiSpansBinaryRelationsAndLabeledPartitions,
10
  TextDocumentWithLabeledSpansAndBinaryRelations,
11
  TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions,
12
  )
13
+ from pytorch_ie.core import Document
14
 
15
  from pie_datasets.builders import BratBuilder, BratConfig
16
+ from pie_datasets.builders.brat import BratDocument, BratDocumentWithMergedSpans
17
+ from pie_datasets.core.dataset import DocumentConvertersType
18
  from pie_datasets.document.processing import Caster, Pipeline
19
 
20
  URL = "http://data.dws.informatik.uni-mannheim.de/sci-arg/compiled_corpus.zip"
21
  SPLIT_PATHS = {"train": "compiled_corpus"}
22
 
23
 
24
+ def get_common_converter_pipeline_steps(target_document_type: type[Document]) -> dict:
25
  return dict(
26
  cast=Caster(
27
  document_type=target_document_type,
 
35
  )
36
 
37
 
38
+ def get_common_converter_pipeline_steps_with_resolve_parts_of_same(
39
+ target_document_type: type[Document],
40
+ ) -> dict:
41
+ return dict(
42
+ cast=Caster(
43
+ document_type=target_document_type,
44
+ field_mapping={"spans": "labeled_multi_spans", "relations": "binary_relations"},
45
+ ),
46
+ trim_adus=TextSpanTrimmer(layer="labeled_multi_spans"),
47
+ sort_symmetric_relation_arguments=RelationArgumentSorter(
48
+ relation_layer="binary_relations",
49
+ label_whitelist=["semantically_same"],
50
+ ),
51
+ )
52
+
53
+
54
+ class SciArgConfig(BratConfig):
55
+ def __init__(
56
+ self,
57
+ name: str,
58
+ resolve_parts_of_same: bool = False,
59
+ **kwargs,
60
+ ):
61
+ super().__init__(name=name, merge_fragmented_spans=True, **kwargs)
62
+ self.resolve_parts_of_same = resolve_parts_of_same
63
+
64
+
65
  class SciArg(BratBuilder):
66
  BASE_DATASET_PATH = "DFKI-SLT/brat"
67
  BASE_DATASET_REVISION = "844de61e8a00dc6a93fc29dc185f6e617131fbf1"
 
70
  # The span fragments in SciArg come just from the new line splits, so we can merge them.
71
  # Actual span fragments are annotated via "parts_of_same" relations.
72
  BUILDER_CONFIGS = [
73
+ SciArgConfig(name=BratBuilder.DEFAULT_CONFIG_NAME),
74
+ SciArgConfig(name="resolve_parts_of_same", resolve_parts_of_same=True),
75
  ]
76
  DOCUMENT_TYPES = {
77
  BratBuilder.DEFAULT_CONFIG_NAME: BratDocumentWithMergedSpans,
78
+ "resolve_parts_of_same": BratDocument,
79
  }
80
 
81
  # we need to add None to the list of dataset variants to support the default dataset variant
82
  BASE_BUILDER_KWARGS_DICT = {
83
  dataset_variant: {"url": URL, "split_paths": SPLIT_PATHS}
84
+ for dataset_variant in ["default", "resolve_parts_of_same", None]
85
  }
86
 
87
+ def _generate_document(self, example, **kwargs):
88
+ document = super()._generate_document(example, **kwargs)
89
+ if self.config.resolve_parts_of_same:
90
+ document = SpansViaRelationMerger(
91
+ relation_layer="relations",
92
+ link_relation_label="parts_of_same",
93
+ create_multi_spans=True,
94
+ result_document_type=BratDocument,
95
+ result_field_mapping={"spans": "spans", "relations": "relations"},
96
+ )(document)
97
+ return document
98
+
99
+ @property
100
+ def document_converters(self) -> DocumentConvertersType:
101
+ regex_partitioner = RegexPartitioner(
102
+ partition_layer_name="labeled_partitions",
103
+ pattern="<([^>/]+)>.*</\\1>",
104
+ label_group_id=1,
105
+ label_whitelist=["Title", "Abstract", "H1"],
106
+ skip_initial_partition=True,
107
+ strip_whitespace=True,
108
+ )
109
+ if not self.config.resolve_parts_of_same:
110
+ return {
111
+ TextDocumentWithLabeledSpansAndBinaryRelations: Pipeline(
112
+ **get_common_converter_pipeline_steps(
113
+ TextDocumentWithLabeledSpansAndBinaryRelations
114
+ )
115
+ ),
116
+ TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions: Pipeline(
117
+ **get_common_converter_pipeline_steps(
118
+ TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions
119
+ ),
120
+ add_partitions=regex_partitioner,
121
+ ),
122
+ }
123
+ else:
124
+ return {
125
+ # TextDocumentWithLabeledSpansAndBinaryRelations: Pipeline(
126
+ # **get_common_converter_pipeline_steps_with_resolve_parts_of_same(
127
+ # TextDocumentWithLabeledSpansAndBinaryRelations
128
+ # )
129
+ # ),
130
+ # TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions: Pipeline(
131
+ # **get_common_converter_pipeline_steps_with_resolve_parts_of_same(
132
+ # TextDocumentWithLabeledSpansBinaryRelationsAndLabeledPartitions
133
+ # ),
134
+ # add_partitions=regex_partitioner,
135
+ # ),
136
+ TextDocumentWithLabeledMultiSpansAndBinaryRelations: Pipeline(
137
+ **get_common_converter_pipeline_steps_with_resolve_parts_of_same(
138
+ TextDocumentWithLabeledMultiSpansAndBinaryRelations
139
+ )
140
+ ),
141
+ TextDocumentWithLabeledMultiSpansBinaryRelationsAndLabeledPartitions: Pipeline(
142
+ **get_common_converter_pipeline_steps_with_resolve_parts_of_same(
143
+ TextDocumentWithLabeledMultiSpansBinaryRelationsAndLabeledPartitions
144
+ ),
145
+ add_partitions=regex_partitioner,
146
+ ),
147
+ }