Dataset Card for GEM/SIMPITIKI

Link to Main Data Card

You can find the main data card on the GEM Website .

Dataset Summary

SIMPITIKI is an Italian Simplification dataset. Its examples were selected from Italian Wikipedia such that their editing tracking descriptions contain any of the words "Simplified"/"Simplify"/"Simplification".

You can load the dataset via:

import datasets
data = datasets.load_dataset('GEM/SIMPITIKI')

The data loader can be found here .

website

Github

paper

Website

authors

Sara Tonelli (Fondazione Bruno Kessler), Alessio Palmero Aprosio (Fondazione Bruno Kessler), Francesca Saltori (Fondazione Bruno Kessler)

Dataset Overview

Where to find the Data and its Documentation

Webpage

Github

Download

Github

Paper

Website

BibTex

@article{tonelli2016simpitiki,
  title={SIMPITIKI: a Simplification corpus for Italian},
  author={Tonelli, Sara and Aprosio, Alessio Palmero and Saltori, Francesca},
  journal={Proceedings of CLiC-it},
  year={2016}
}

Contact Name

Sara Tonelli

Contact Email

satonelli@fbk.eu

Has a Leaderboard?

no

Languages and Intended Use

Multilingual?

no

Covered Dialects

None

Covered Languages

Italian

License

cc-by-4.0: Creative Commons Attribution 4.0 International

Intended Use

The purpose of the dataset is to train NLG models to simplify complex text by learning different types of transformations (verb to noun, noun to verbs, deletion, insertion, etc)

Primary Task

Simplification

Communicative Goal

This dataset aims to enhance research in text simplification in Italian language with different text transformations.

Credit

Curation Organization Type(s)

academic , independent

Curation Organization(s)

Fondazione Bruno Kessler (FBK)

Dataset Creators

Sara Tonelli (Fondazione Bruno Kessler), Alessio Palmero Aprosio (Fondazione Bruno Kessler), Francesca Saltori (Fondazione Bruno Kessler)

Funding

EU Horizon 2020 Programme via the SIMPATICO Project (H2020-EURO-6-2015, n. 692819)

Who added the Dataset to GEM?

Sebastien Montella (Orange Labs), Vipul Raheja (Grammarly Inc.)

Dataset Structure

Data Fields

Each sample comes with the following fields:

• gem_id (string): Unique sample ID - text (string): The raw text to be simplified - simplified_text (string): The simplified version of "text" field - transformation_type (string): Nature of transformation applied to raw text in order to simplify it. - source_dataset (string): Initial dataset source of sample. Values: 'itwiki' (for Italian Wikipedia) or 'tn' (manually annotated administrative documents from the Municipality of Trento, Italy)

Reason for Structure

The dataset is organized as a pairs where the raw text (input) is associated with its simplified text (output). The editing transformation and the source dataset of each sample is also provided for advanced analysis.

How were labels chosen?

SIMPITIKI dataset selects documents from Italian Wikipedia such that their editing tracking descriptions contain any of the words "Simplified"/"Simplify"/"Simplification". For the Public Administration domain of the documents of the Municipality of Trento (Italy)

Example Instance

{"transformation_id": 31, "transformation_type": "Transformation - Lexical Substitution (word level)", "source_dataset": "tn", "text": "- assenza per <del>e</del>si<del>genze</del> particolari attestate da relazione dei servizi sociali;", "simplified_text": "- assenza per <ins>bi</ins>s<ins>ogn</ins>i particolari attestati da relazione dei servizi sociali;"}

Data Splits

Several splits are proposed to train models on different configurations:

-"train": Training samples randomly selected from initial corpus. 816 training samples. -"validation": Validating samples randomly selected from initial corpus. 174 validating samples. -"test": Testing samples randomly selected from initial corpus. 176 validating samples. -"challenge_seen_transformations_train": This training challenge split includes specific transformations to simplify the raw text. Precisely, transformations are "Split", "Merge", "Reordering", "Insert - Verb", "Insert - Other", "Delete - Verb", "Delete - Other", "Transformation - Lexical Substitution (word level)", "Transformation - Anaphoric replacement", "Transformation - Noun to Verb", "Transformation - Verbal Features". 562 training samples. -"challenge_seen_transformations_val": This validating challenge split includes same transformations than the ones observed in training. Precisely, transformations are "Split", "Merge", "Reordering", "Insert - Verb", "Insert - Other", "Delete - Verb", "Delete - Other", "Transformation - Lexical Substitution (word level)", "Transformation - Anaphoric replacement", "Transformation - Noun to Verb", "Transformation - Verbal Features". 121 validating samples. -"challenge_seen_transformations_test": This testing challenge split includes same transformations than the ones observed in training. Precisely, transformations are "Split", "Merge", "Reordering", "Insert - Verb", "Insert - Other", "Delete - Verb", "Delete - Other", "Transformation - Lexical Substitution (word level)", "Transformation - Anaphoric replacement", "Transformation - Noun to Verb", "Transformation - Verbal Features". 127 testing samples. -"challenge_unseen_transformations_test" : "Insert - Subject", "Delete - Subject", "Transformation - Lexical Substitution (phrase level)", "Transformation - Verb to Noun (nominalization)", "Transformation - Verbal Voice". 356 testing samples. -"challenge_itwiki_train": This challenge split includes random samples from the Italian Wikipedia as source dataset. 402 training samples. -"challenge_itwiki_val": This validating challenge split includes random samples from the Italian Wikipedia as source dataset. 86 validating samples. -"challenge_itwiki_test": This testing challenge split includes random samples from the Italian Wikipedia as source dataset. 87 testing samples. -"challenge_tn_test": This testing challenge split includes all samples from the Municipality of Trento administrative documents ('tn') as source dataset. 591 testing samples.

Splitting Criteria

The training ratio is set to 0.7. The validation and test somehow equally divide the remaining 30% of the dataset.

Dataset in GEM

Rationale for Inclusion in GEM

Why is the Dataset in GEM?

This dataset promotes Simplification task for Italian language.

Similar Datasets

no

Ability that the Dataset measures

Models can be evaluated if they can simplify text regarding different simplification transformations.

GEM-Specific Curation

Modificatied for GEM?

yes

Additional Splits?

yes

Split Information

The SIMPITIKI dataset provides a single file. Several splits are proposed to train models on different configurations: -"train": Training samples randomly selected from initial corpus. 816 training samples. -"validation": Validating samples randomly selected from initial corpus. 174 validating samples. -"test": Testing samples randomly selected from initial corpus. 176 validating samples. -"challenge_seen_transformations_train": This training challenge split includes specific transformations to simplify the raw text. Precisely, transformations are "Split", "Merge", "Reordering", "Insert - Verb", "Insert - Other", "Delete - Verb", "Delete - Other", "Transformation - Lexical Substitution (word level)", "Transformation - Anaphoric replacement", "Transformation - Noun to Verb", "Transformation - Verbal Features". 562 training samples. -"challenge_seen_transformations_val": This validating challenge split includes same transformations than the ones observed in training. Precisely, transformations are "Split", "Merge", "Reordering", "Insert - Verb", "Insert - Other", "Delete - Verb", "Delete - Other", "Transformation - Lexical Substitution (word level)", "Transformation - Anaphoric replacement", "Transformation - Noun to Verb", "Transformation - Verbal Features". 121 validating samples. -"challenge_seen_transformations_test": This testing challenge split includes same transformations than the ones observed in training. Precisely, transformations are "Split", "Merge", "Reordering", "Insert - Verb", "Insert - Other", "Delete - Verb", "Delete - Other", "Transformation - Lexical Substitution (word level)", "Transformation - Anaphoric replacement", "Transformation - Noun to Verb", "Transformation - Verbal Features". 127 testing samples. -"challenge_unseen_transformations_test" : "Insert - Subject", "Delete - Subject", "Transformation - Lexical Substitution (phrase level)", "Transformation - Verb to Noun (nominalization)", "Transformation - Verbal Voice". 356 testing samples. -"challenge_itwiki_train": This challenge split includes random samples from the Italian Wikipedia as source dataset. 402 training samples. -"challenge_itwiki_val": This validating challenge split includes random samples from the Italian Wikipedia as source dataset. 86 validating samples. -"challenge_itwiki_test": This testing challenge split includes random samples from the Italian Wikipedia as source dataset. 87 testing samples. -"challenge_tn_test": This testing challenge split includes all samples from the Municipality of Trento administrative documents ('tn') as source dataset. 591 testing samples.

Split Motivation

The splits allows to investigate the generalization of models regarding editing/transformations ("challenge_seen_transformations_test" / "challenge_unseen_transformations_test") and for transfer learning to different domain ("challenge_tn_test")

Getting Started with the Task

Pointers to Resources

• Coster and Kauchak, Simple English Wikipedia: A New Text Simplification Task, Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 665–669, Portland, Oregon, June 19-24, 2011
• Xu et al, Optimizing Statistical Machine Translation for Text Simplification, Transactions of the Association for Computational Linguistics, vol. 4, pp. 401–415, 2016
• Aprosio et al, Neural Text Simplification in Low-Resource Conditions Using Weak Supervision, Proceedings of the Workshop on Methods for Optimizing and Evaluating Neural Language Generation (NeuralGen), pages 37–44, Minneapolis, Minnesota, USA, June 6, 2019

Technical Terms

Simplification: Process that consists in transforming an input text to its simplified version.

Previous Results

Previous Results

Measured Model Abilities

The splits allows to investigate the generalization of models regarding editing/transformations ("challenge_seen_transformations_test" / "challenge_unseen_transformations_test") and for transfer learning to different domain ("challenge_tn_test")

Metrics

BLEU , Other: Other Metrics

Other Metrics

FKBLEU ( https://aclanthology.org/Q16-1029.pdf ): Combines Flesch-Kincaid Index and iBLEU metrics. SARI ( https://aclanthology.org/Q16-1029.pdf ): Compares system output against references and against the input sentence. It explicitly measures the goodness of words that are added, deleted and kept by the systems Word-level F1

Previous results available?

no

Dataset Curation

Original Curation

Original Curation Rationale

Most of the resources for Text Simplification are in English. To stimulate research to different languages, SIMPITIKI proposes an Italian corpus with Complex-Simple sentence pairs.

Communicative Goal

Text simplification allows a smooth reading of text to enhance understanding.

Sourced from Different Sources

yes

Source Details

Italian Wikipedia (Manually) Annotated administrative documents from the Municipality of Trento, Italy

Language Data

How was Language Data Obtained?

Found

Where was it found?

Single website , Offline media collection

Language Producers

SIMPITIKI is a combination of documents from Italian Wikipedia and from the Municipality of Trento, Italy.

Topics Covered

Samples from documents from the Municipality of Trento corpus are in the administrative domain.

Data Validation

validated by data curator

Was Data Filtered?

not filtered

Structured Annotations

Additional Annotations?

crowd-sourced

Number of Raters

unknown

Rater Qualifications

Native speaker

Raters per Training Example

0

Raters per Test Example

0

Annotation Service?

unknown

Annotation Values

Annotators specified any of the tags as designed by Brunato et al. ( https://aclanthology.org/W15-1604/ ): -Split: Splitting a clause into two clauses. -Merge: Merge two or more clauses together. -Reordering: Word order changes. -Insert: Insertion of words or phrases that provide supportive information to the original sentence -Delete: dropping redundant information. -Transformation: Modification which can affect the sentence at the lexical, morpho-syntactic and syntactic level: Lexical substitution (word level) / Lexical substitution (phrase level) / Anaphoric replacement / Noun to Verb / Verb to Noun / Verbal voice / Verbal features/ morpho–syntactic and syntactic level, also giving rise to overlapping phenomena

Any Quality Control?

unknown

Consent

Any Consent Policy?

no

Justification for Using the Data

The dataset is available online under the CC-BY 4.0 license.

Private Identifying Information (PII)

Contains PII?

likely

Categories of PII

generic PII

Any PII Identification?

no identification

Maintenance

Any Maintenance Plan?

no

Broader Social Context

Previous Work on the Social Impact of the Dataset

Usage of Models based on the Data

no

Impact on Under-Served Communities

Addresses needs of underserved Communities?

yes

Details on how Dataset Addresses the Needs

The creator of SIMPITIKI wants to promote text simplification for Italian because few resources are available in other languages than English.

Discussion of Biases

Any Documented Social Biases?

unsure

Considerations for Using the Data

PII Risks and Liability

Licenses

Copyright Restrictions on the Dataset

research use only

Copyright Restrictions on the Language Data

research use only

Known Technical Limitations

Discouraged Use Cases

The risk of surface-based metrics (BLEU, chrf++, etc) for this task is that semantic adequacy is not respected when simplifying the input document.