Hugging Face's logo

Datasets:
patriziobellan
/
PET

Tasks:
Token Classification
Languages:
English
Size:
n<1K
ArXiv:
Tags:
Business Process Management
NLP
ML
DL
License:
Dataset card Files Community
6
PET / README.md
patriziobellan's picture
patriziobellan
Update README.md
9742bba
preview code
|
raw
history blame
No virus
8.32 kB
metadata
annotations_creators:
  - expert-generated
language_creators:
  - expert-generated
languages:
  - en-US
licenses:
  - mit
multilinguality:
  - monolingual
pretty_name: PET
size_categories:
  - 1K<n<10K
source_datasets:
  - Friedrich et al. original dataset
task_categories:
  - token-classification
task_ids:
  - named-entity-recognition

Dataset Card for PET

Table of Contents

Dataset Description

Dataset Summary

Abstract. Although there is a long tradition of work in NLP on extracting entities and relations from text, to date there exists little work on the acquisition of business processes from unstructured data such as textual corpora of process descriptions. With this work we aim at filling this gap and establishing the first steps towards bridging data-driven information extraction methodologies from Natural Language Processing and the model-based formalization that is aimed from Business Process Management. For this, we develop the first corpus of business process descriptions annotated with activities, actors, activity data, gateways and their conditions. We present our new resource to benchmark the difficulty and challenges of business process extraction from text.

Supported Tasks and Leaderboards

Named Entity Recognition

Languages

English

Dataset Structure

Test set to beanchmark Business Process Extraction from Text approaches.

Data Instances

For each instance, there is a document name representing the name of the document of the Friedrich et al. dataset, an integer representing the number of the sentence, a list of tokens representing the words of the sentence instance, and a list of ner tags (in IOB2 format) representing the annotation of process elements of the sentence.

{
  "document name":"doc-1.1",
  "sentence-ID":1,
  "tokens":["Whenever","the","sales","department","receives","an","order",",","a","new","process","instance","is","created","."],
  "ner-tags":["O","B-Actor","I-Actor","I-Actor","B-Activity","B-Activity Data","I-Activity Data","O","O","O","O","O","O","O","O"]
}

Data Fields

  • document name: a string used to represent the name of the document.
  • sentence-ID: an integer (starting from 0) representing the number of the sentence within the document.
  • tokens: a list of string representing the words of the sentence
  • ner-tags: a list of string representing the annotation for each word.

The allowed ner-tags are:

  • "O":
  • "B-Actor",
  • "I-Actor",
  • "B-Activity",
  • "I-Activity",
  • "B-Activity Data",
  • "I-Activity Data",
  • "B-Further Specification",
  • "I-Further Specification",
  • "B-XOR Gateway",
  • "I-XOR Gateway",
  • "B-Condition Specification",
  • "I-Condition Specification",
  • "B-AND Gateway",
  • "I-AND Gateway",

Data Splits

The data was not splited. It contains the test set only.

Dataset Creation

The dataset construction process has been split in five main phases:

  1. Text pre-processing. As the first operation, we check the content of each document and we tokenized it. This initial check was necessary since some of the original texts were automatically translated into English by the authors of the dataset. The translations were never validated, indeed, several errors have been found and fixed.

  2. Text Annotation. Each text has been annotated by using the guidelines (https://pdi.fbk.eu/pet/annotation-guidelines-for-process-description.pdf). The team was composed by five annotators with high expertise in BPMN. Each document has been assigned to three experts that were in change of identifying all the elements and flows with each document. In this phase, we used the the Inception tool to support annotators.

  3. Automatic annotation fixing. After the second phase, we ran an automatic procedure relying on a rule-based script to automatically fix annotations that were not compliant with the guidelines. For example, if a modal verb was erroneously included in the annotation of an Activity, the procedure removed it from the annotation. Another example is the missing of the article within an annotation related to an Actor. In this case, the script included it in the annotation. This phase allowed to remove possible annotation errors and to obtain annotations compliant with the guidelines.

  4. Agreement Computation. Here, we computed, on the annotation provided by the experts, the agreement scores for each process element and for each relation between process elements pair adopting the methodology proposed in (George Hripcsak and Adam S. Rothschild. Technical brief: Agreement, the f-measure, and reliability in information retrieval. J. Am. Medical Informatics Assoc., 12(3):296–298, 2005.). We measured the agreement in terms of the F1 measure because, besides being straightforward to calculate, it is directly interpretable. Note that chance-corrected measures like κ approach the F1-measure as the number of cases that raters agree are negative grows. By following such a methodology, an annotation was considered in agreement among the experts if and only if they capture the same span of words and they assign the same process element tag to the annotation.

  5. Reconciliation. The last phase consisted of the mitigation of disagreements within the annotations provided by the experts. The aim of this phase is to obtain a shared and agreed set of gold standard annotations on each text for both entities and relations. Such entities also enable the generation of the related full-connected process model flow that can be rendered by using, but not limited to, a BPMN diagram. During this last phase, among the 47 documents originally included into the dataset, 2 of them were discarded. These texts were not fully annotated by the annotators since they were not be able to completely understand which process elements were actually included in some specific parts of the text. For this reason, the final size of the dataset is 45 textual descriptions of the corresponding process models together with their annotations.

Curation Rationale

[More Information Needed]

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

English

Annotations

Annotation process

You can read about the annotation process in the original paper https://arxiv.org/abs/2203.04860

Who are the annotators?

Expert Annotators

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

  • Patrizio Bellan (
  • Mauro Dragoni
  • Chiara Ghidini
  • Han van der Aa
  • Simone Ponzetto

Licensing Information

MIT

Citation Information

Bellan, Patrizio et al. “PET: A new Dataset for Process Extraction from Natural Language Text.” ArXiv abs/2203.04860 (2022)

Contributions

Thanks to Patrizio Bellan for adding this dataset.