Datasets:
what's up with the prefixes?
I am not inclined to believe that someone wrote on rivlin's wall:
ו הצלחה ב בחירות ל ה נשיאות
this is clearly the fault of some mechanism that doesn't understand hebrew prefixes.
is there a reason that the data is processed to begin with? I'd imagine training on the original sentence would be better.
Hi! We’re not the authors of this dataset, and we don’t know its full details or Hebrew. This is a re-uploaded version of https://huggingface.co/datasets/omilab/hebrew_sentiment. You can find more information in the source paper: https://aclanthology.org/C18-1190/
thanks for the detailed answer!
looking at the source datasets here: https://github.com/omilab/Neural-Sentiment-Analyzer-for-Modern-Hebrew/tree/master/data
it seems like there is a morph dataset and a token dataset, out of which the token dataset is written properly and the morph dataset is processed (post-tokenization?)
I suggest that the token dataset is the better one for assessing a model's understanding of hebrew.
Great feedback to have,
@lior-at-optune
. We would love a fix for the task on MTEB as opposed to here. We typically version the dataset (to avoid breaking changes in benchmarks), so we would have to make mteb/HebrewSentimentAnalysis-v2