Update README.md

README.md

@@ -145,10 +145,13 @@ It also contains two holdout subsets of 359 identical sentence triplets and 359
 ### 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, 1) and 1,355 data augmentation triplets (Document, Unrelated Document, 0) (See the sanity checks in our [article](https://www.frontiersin.org/articles/10.3389/frai.2023.1223924/full)).
+- `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](https://www.frontiersin.org/articles/10.3389/frai.2023.1223924/full)).
 - `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
+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.
+
 ### Data Fields
 
 - `original`: an original sentence from the source datasets.