RO-Sentiment

This model is a fine-tuned version of readerbench/RoBERT-base on the Decathlon reviews and Cinemagia reviews dataset. It achieves the following results on the evaluation set:

• Loss: 0.3923
• Accuracy: 0.8307
• Precision: 0.8366
• Recall: 0.8959
• F1: 0.8652
• F1 Weighted: 0.8287

Output labels:

• LABEL_0 = Negative Sentiment
• LABEL_1 = Positive Sentiment

Evaluation on other datasets

SENT_RO

	precision	recall	f1-score	support
Negative (0)	0.79	0.83	0.81	11,675
Positive (1)	0.88	0.85	0.87	17,271
Accuracy			0.85	28,946
Macro Avg	0.84	0.84	0.84	28,946
Weighted Avg	0.85	0.85	0.85	28,946

LaRoSeDa

	precision	recall	f1-score	support
Negative (0)	0.79	0.94	0.86	7,500
Positive (1)	0.93	0.75	0.83	7,500
Accuracy			0.85	15,000
Macro Avg	0.86	0.85	0.84	15,000
Weighted Avg	0.86	0.85	0.84	15,000

Model description

Finetuned Romanian BERT model for sentiment classification.

Trained on a mix of product reviews from Decathlon retailer website and movie reviews from cinemagia.

Intended uses & limitations

Sentiment classification for Romanian Language.

Biased towards Product reviews.

There is no "neutral" sentiment label.

Training and evaluation data

Trained on:

• Decathlon Dataset available on request

• Cinemagia Movie reviews public on kaggle Link

Evaluated on

• Holdout data from training dataset
• RO_SENT Dataset
• LaROSeDa Dataset

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 6e-05
• train_batch_size: 64
• eval_batch_size: 128
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.2
• num_epochs: 10 (Early stop epoch 3, best epoch 2)

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Precision	Recall	F1	F1 Weighted
0.4198	1.0	1629	0.3983	0.8377	0.8791	0.8721	0.8756	0.8380
0.3861	2.0	3258	0.4312	0.8429	0.8963	0.8665	0.8812	0.8442
0.3189	3.0	4887	0.3923	0.8307	0.8366	0.8959	0.8652	0.8287

Framework versions

• Transformers 4.31.0
• Pytorch 2.0.1+cu118
• Datasets 2.14.3
• Tokenizers 0.13.3