Dataset Card for "Continuous Scale Meaning Dataset" (CSMD)

CSMD was created for MeaningBERT: Assessing Meaning Preservation Between Sentences.

It contains 1,355 English text simplification meaning preservation annotations. Meaning preservation measures how well the meaning of the output text corresponds to the meaning of the source (Saggion, 2017).

The annotations were taken from the following four datasets:

• ASSET
• QuestEVal,
• SimpDa_2022 and,
• Simplicity-DA.

It contains a data augmentation subset of 1,355 identical sentence triplets and 1,355 unrelated sentence triplets (See the "Sanity Checks" section (3.3.) in our article).

It also contains two holdout subsets of 359 identical sentence triplets and 359 unrelated sentence triples (See the "MeaningBERT" section (3.4.) in our article).

Dataset Structure

Data Instances

• Meaning configuration: an instance consists of 1,355 meaning preservation triplets (Document, simplification, label).
• meaning_with_data_augmentation configuration: an instance consists of 1,355 meaning preservation triplets (Document, simplification, label) along with 1,355 data augmentation triplets (Document, Document, 100) and 1,355 data augmentation triplets (Document, Unrelated Document, 0) (See the sanity checks in our article).
• meaning_holdout_identical configuration: an instance consists of 359 meaning holdout preservation identical triplets (Document, Document, 1) based on the ASSET Simplification dataset.
• meaning_holdout_unrelated configuration: an instance consists of 359 meaning holdout preservation unrelated triplets (Document, Unrelated Document, 0) based on the ASSET Simplification dataset.

About the Data Augmentation

Unrelated Sentence

We have changed the data augmentation approach for the unrelated sentence. Instead of generating noisy sentences using an LLM, for each of the 1,355 sentences, we sample a sentence in the unlabeled sentence in ASSET (non included in the holdout nor the labelled sentence). We compute the Rouge1, Rouge2, RougeL and bleu scores to validate that the sentences are unrelated in terms of vocabulary. Namely, each metric score is below 0.20 or 20 for Bleu for all pairs. If a pair achieves a higher value, we select another sentence from ASSET to create a pair and reapply the test until a pair achieves a score below 0.20/20.

Commutative Property

Since meaning preservation is a commutative function, i.e., Meaning(Sent_a, Sent_b) = Meaning(Sent_b, Sent_a), we also include the commutative version of the triplet in the data augmentation version of the dataset for sentences that are not identical.

Data Fields

• original: an original sentence from the source datasets.
• simplification: a simplification of the original obtained by an automated system or a human.
• label: a meaning preservation rating between 0 and 100.

Data Splits

The split statistics of CSMD are given below.

	Train	Dev	Test	Total
Meaning	853	95	407	1,355
Meaning With Data Augmentation	2,560	285	1,220	4,065
Meaning Holdout Identical	NA	NA	359	359
Meaning Holdout Unrelated	NA	NA	359	359

All the splits are randomly split using a 60-10-30 split with the seed 42.

Citation Information

@ARTICLE{10.3389/frai.2023.1223924,
AUTHOR={Beauchemin, David and Saggion, Horacio and Khoury, Richard},   
TITLE={{MeaningBERT: Assessing Meaning Preservation Between Sentences}},      
JOURNAL={Frontiers in Artificial Intelligence},      
VOLUME={6},           
YEAR={2023},      
URL={https://www.frontiersin.org/articles/10.3389/frai.2023.1223924},       
DOI={10.3389/frai.2023.1223924},        
ISSN={2624-8212},   
}