language: fr
license: mit
datasets:
- amazon_reviews_multi
- allocine
widget:
- text: Je pensais lire un livre nul, mais finalement je l'ai trouvé super !
- text: >-
Cette banque est très bien, mais elle n'offre pas les services de
paiements sans contact.
- text: >-
Cette banque est très bien et elle offre en plus les services de paiements
sans contact.
DistilCamemBERT-Sentiment
We present DistilCamemBERT-Sentiment which is DistilCamemBERT fine tuned for the sentiment analysis task for the French language. This model is constructed over 2 datasets: Amazon Reviews and Allociné.fr in order to minimize the bias. Indeed, Amazon reviews are very similar in the messages and relatively shorts, contrary to Allociné critics which are long and rich texts.
This modelization is close to tblard/tf-allocine based on CamemBERT model. The problem of the modelizations based on CamemBERT is at the scaling moment, for the production phase for example. Indeed, inference cost can be a technological issue. To counteract this effect, we propose this modelization which divides the inference time by 2 with the same consumption power thanks to DistilCamemBERT.
Dataset
The dataset is composed of XXX,XXX reviews for training and X,XXX reviews for the test coming from Amazon, and respectively XXX,XXX and X,XXX critics from Allocine website. The dataset is labeled into 5 categories: * 1 star: represents a very bad appreciation, * 2 stars: bad appreciation, * 3 stars: neutral appreciation, * 4 stars: good appreciation, * 5 stars: very good appreciation. Evaluation results
Benchmark
This model is compared to 3 reference models (see below). As each model doesn't have the same definition of targets, we detail the performance measure used for each of them. For the mean inference time measure, an AMD Ryzen 5 4500U @ 2.3GHz with 6 cores was used.
bert-base-multilingual-uncased-sentiment
nlptown/bert-base-multilingual-uncased-sentiment is based on BERT model in the multilingual and uncased version. This sentiment analyzer is trained on Amazon reviews similarly to our model, hence the targets and their definitions are the same. In order to be robust to +/-1 star estimation errors, we will take the following definition as a performance measure: where $\mathcal{O}$ is the test set of the observations, $p_l\in{0,1}$ is equal to 1 for the true label and $\hat{p}_l$ is the estimated probability for the l-th label.
tf-allociné and barthez-sentiment-classification
tblard/tf-allocine and moussaKam/barthez-sentiment-classification use the same bi-class definition between them. To bring this back to a two-class problem, we will only consider the "1 star" and "2 stars" labels for the "negative" sentiments and "4 stars" and "5 stars" for "positive" sentiments. We exclude the "3 stars" which can be interpreted as a "neutral" class. In this context, the problem of +/-1 star estimation errors disappears. Then we use the classical accuracy definition.
How to use DistilCamemBERT-Sentiment
from transformers import pipeline
analyzer = pipeline(
task='text-classification',
model="cmarkea/distilcamembert-base-sentiment",
tokenizer="cmarkea/distilcamembert-base-sentiment"
)
result = analyzer(
"J'aime me promener en forêt même si ça me donne mal aux pieds.",
return_all_scores=True
)
result
[{'label': '1 star',
'score': 0.047529436647892},
{'label': '2 stars',
'score': 0.14150355756282806},
{'label': '3 stars',
'score': 0.3586442470550537},
{'label': '4 stars',
'score': 0.3181498646736145},
{'label': '5 stars',
'score': 0.13417290151119232}]