PIE Dataset Card for "argmicro"

This is a PyTorch-IE wrapper for the ArgMicro Huggingface dataset loading script.

Usage

from pie_datasets import load_dataset
from pytorch_ie.documents import TextDocumentWithLabeledSpansAndBinaryRelations

# load English variant
dataset = load_dataset("pie/argmicro", name="en")

# if required, normalize the document type (see section Document Converters below)
dataset_converted = dataset.to_document_type(TextDocumentWithLabeledSpansAndBinaryRelations)
assert isinstance(dataset_converted["train"][0], TextDocumentWithLabeledSpansAndBinaryRelations)

# get first relation in the first document
doc = dataset_converted["train"][0]
print(doc.binary_relations[0])
# BinaryRelation(head=LabeledSpan(start=0, end=81, label='opp', score=1.0), tail=LabeledSpan(start=326, end=402, label='pro', score=1.0), label='reb', score=1.0)
print(doc.binary_relations[0].resolve())
# ('reb', (('opp', "Yes, it's annoying and cumbersome to separate your rubbish properly all the time."), ('pro', 'We Berliners should take the chance and become pioneers in waste separation!')))

Dataset Variants

The dataset contains two BuilderConfig's:

• de: with the original texts collection in German
• en: with the English-translated texts

Data Schema

The document type for this dataset is ArgMicroDocument which defines the following data fields:

• text (str)
• id (str, optional)
• topic_id (str, optional)
• metadata (dictionary, optional)

and the following annotation layers:

• stance (annotation type: Label)
  • description: A document may contain one of these stance labels: pro, con, unclear, or no label when it is undefined (see here for reference).
• edus (annotation type: Span, target: text)
• adus (annotation type: LabeledAnnotationCollection, target: edus)
  • description: each element of adus may consist of several entries from edus, so we require LabeledAnnotationCollection as annotation type. This is originally indicated by seg edges in the data.
  • LabeledAnnotationCollection has the following fields:
    • annotations (annotation type: Span, target: text)
    • label (str, optional), values: opp, pro (see here)
• relations (annotation type: MultiRelation, target: adus)
  • description: Undercut (und) relations originally target other relations (i.e. edges), but we let them target the head of the targeted relation instead. The original state can be deterministically reconstructed by taking the label into account. Furthermore, the head of additional source (add) relations are integrated into the head of the target relation (note that this propagates along und relations). We model this with MultiRelations whose head and tail are of type LabeledAnnotationCollection.
  • MultiRelation has the following fields:
    • head (tuple, annotation type: LabeledAnnotationCollection, target: adus)
    • tail (tuple, annotation type: LabeledAnnotationCollection, target: adus)
    • label (str, optional), values: sup, exa, reb, und (see here for reference, but note that helper relations seg and add are not there anymore, see above).

See here for the annotation type definitions.

Document Converters

The dataset provides document converters for the following target document types:

• pytorch_ie.documents.TextDocumentWithLabeledSpansAndBinaryRelations
  • LabeledSpans, converted from ArgMicroDocument's adus
    • labels: opp, pro
    • if an ADU contains multiple spans (i.e. EDUs), we take the start of the first EDU and the end of the last EDU as the boundaries of the new LabeledSpan. We also raise exceptions if any newly created LabeledSpans overlap.
  • BinraryRelations, converted from ArgMicroDocument's relations
    • labels: sup, reb, und, joint, exa
    • if the head or tail consists of multiple adus, then we build BinaryRelations with all head-tail combinations and take the label from the original relation. Then, we build BinaryRelations' with label joint between each component that previously belongs to the same head or tail, respectively.
  • metadata, we keep the ArgMicroDocument's metadata, but stance and topic_id.

See here for the document type definitions.

Collected Statistics after Document Conversion

We use the script evaluate_documents.py from PyTorch-IE-Hydra-Template to generate these statistics. After checking out that code, the statistics and plots can be generated by the command:

python src/evaluate_documents.py dataset=argmicro_base metric=METRIC

where a METRIC is called according to the available metric configs in config/metric/METRIC (see metrics).

This also requires to have the following dataset config in configs/dataset/argmicro_base.yaml of this dataset within the repo directory:

_target_: src.utils.execute_pipeline
input:
  _target_: pie_datasets.DatasetDict.load_dataset
  path: pie/argmicro
  revision: 28ef031d2a2c97be7e9ed360e1a5b20bd55b57b2
  name: en

For token based metrics, this uses bert-base-uncased from transformer.AutoTokenizer (see AutoTokenizer, and bert-based-uncased to tokenize text in TextDocumentWithLabeledSpansAndBinaryRelations (see document type).

Relation argument (outer) token distance per label

The distance is measured from the first token of the first argumentative unit to the last token of the last unit, a.k.a. outer distance.

We collect the following statistics: number of documents in the split (no. doc), no. of relations (len), mean of token distance (mean), standard deviation of the distance (std), minimum outer distance (min), and maximum outer distance (max). We also present histograms in the collapsible, showing the distribution of these relation distances (x-axis; and unit-counts in y-axis), accordingly.

Command

python src/evaluate_documents.py dataset=argmicro_base metric=relation_argument_token_distances

	len	max	mean	min	std
ALL	1018	127	44.434	14	21.501
exa	18	63	33.556	16	13.056
joint	88	48	30.091	17	9.075
reb	220	127	49.327	16	24.653
sup	562	124	46.534	14	22.079
und	130	84	38.292	17	12.321

Histogram (split: train, 112 documents)

Span lengths (tokens)

The span length is measured from the first token of the first argumentative unit to the last token of the particular unit.

We collect the following statistics: number of documents in the split (no. doc), no. of spans (len), mean of number of tokens in a span (mean), standard deviation of the number of tokens (std), minimum tokens in a span (min), and maximum tokens in a span (max). We also present histograms in the collapsible, showing the distribution of these token-numbers (x-axis; and unit-counts in y-axis), accordingly.

Command

python src/evaluate_documents.py dataset=argmicro_base metric=span_lengths_tokens

statistics	train
no. doc	112
len	576
mean	16.365
std	6.545
min	4
max	41

Histogram (split: train, 112 documents)

Token length (tokens)

The token length is measured from the first token of the document to the last one.

We collect the following statistics: number of documents in the split (no. doc), mean of document token-length (mean), standard deviation of the length (std), minimum number of tokens in a document (min), and maximum number of tokens in a document (max). We also present histograms in the collapsible, showing the distribution of these token lengths (x-axis; and unit-counts in y-axis), accordingly.

Command

python src/evaluate_documents.py dataset=argmicro_base metric=count_text_tokens

statistics	train
no. doc	112
mean	84.161
std	22.596
min	36
max	153

Histogram (split: train, 112 documents)