Datasets:
More information about the dataset
Hi Eduardo,
Thanks for this dataset. I'm thinking of using it for an upcoming NLP training and would need more information about :
- the data source (where does it come from ? What period... ?)
- the sentiment labels (how were they computed ?...)
Thanks in advance for your reply. I'll happily quote your work on the data and can provide a 2K subset translated to English :)
All the best from Paris,
Romain
Hi Romain,
The original dataset is actually in French (https://www.opinion-assurances.fr/). I downloaded it from Kaggle (https://www.kaggle.com/datasets/fedi1996/insurance-reviews-france), and then ran it thru my translation pipeline to have the dataset in the Dutch language. the sentiment labels are 1 for positive and -1 for negative.
Kind Regards
Eduardo
Thanks for your quick reply Eduardo !
I'd suggest to add this info in the dataset card as it could be useful to everybody :)
May I ask what you used for translation and sentiment analysis (a quick look revealed several counter-intuitive labels) ?
Thanks again, cheers
Romain
Filling in the dataset card was on my to-do list but I never got around to actually do it. Regarding the sentiment analysis classification, I suspect it was done by https://www.kaggle.com/fedi1996, since I got the classified dataset from his Kaggle space. And yes you are quite right, there is room for improvement when you start looking deeper at the classification.
If you have the time and patience to curate the dataset further, I'm sure you'll get even better results.
For the translation pipeline I use the Google translate python library googletrans==3.1.0a0 . You need to use that exact version as all other versions throw all kinds of errors.
Cheers
eduardo
Thanks again Eduardo, much appreciated :)
On my side I'm using GPT-3.5 with OpenAI API for translation.
Cheers,
Romain
Yes of course, with the very low price per inference using the chatgpt api, it does make sense to let ChatGPT do the heavy lifting of curating the data.
good luck with the project!
PS.: I'm definitely interested in a cleaner version of the dataset if you care to share it at the end.
Cheers
Eduardo
Thanks again Eduardo !
Happy to share but I'm afraid I won't have much of interest to you - mainly translating from FR to EN, and probably a limited subset only (for some reason I'm having a hard time calling the OpenAI API in a reliable fashion from within our company environment).
Cheers,
Romain