This model is a second fine-tuned version of cardiffnlp/twitter-roberta-base-sentiment, trained on the Sp1786/multiclass-sentiment-analysis-dataset. It builds on the first iteration by incorporating optimized techniques. Specifically, the dataset proportions were adjusted to refine the division of the training, evaluation, and test sets, leading to a more balanced and representative fine-tuning process. Compared to the initial model, this version demonstrates improved performance, with enhanced accuracy and robustness for the task.

It is specifically fine-tuned to analyze user-generated content such as opinions, reviews, comments, and general customer feedback. It is designed for sentiment analysis in the context of understanding public perception, trend analysis, and gathering insights into consumer satisfaction.

Try it out

You can interact with the model directly through the Inference Endpoint:

Full classification example in Pyhton:

python

from transformers import pipeline

pipe = pipeline(model="delarosajav95/tw-roberta-base-sentiment-FT-v2")

inputs = ["The flat is very nice but it's too expensive and the location is very bad.",
  "I loved the music, but the crowd was too rowdy to enjoy it properly.",
  "They believe that I'm stupid and I like waiting for hours in line to buy a simple coffee."
]

result = pipe(inputs, return_all_scores=True)

label_mapping = {"LABEL_0": "Negative", "LABEL_1": "Neutral", "LABEL_2": "Positive"}
for i, predictions in enumerate(result):
  print("==================================")
  print(f"Text {i + 1}: {inputs[i]}")
  for pred in predictions:
    label = label_mapping.get(pred['label'], pred['label'])
    score = pred['score']
    print(f"{label}: {score:.2%}")

Output:

pyhton

==================================
Text 1: The flat is very nice but it's too expensive and the location is very bad.
Negative: 78.54%
Neutral: 20.66%
Positive: 0.80%
==================================
Text 2: I loved the music, but the crowd was too rowdy to enjoy it properly.
Negative: 5.18%
Neutral: 93.34%
Positive: 1.48%
==================================
Text 3: They believe that I'm stupid and I like waiting for hours in line to buy a simple coffee.
Negative: 82.37%
Neutral: 16.85%
Positive: 0.79%

Pipeline API:

pyhton

from transformers import pipeline

url = "delarosajav95/tw-roberta-base-sentiment-FT-v2"

classifier = pipeline("sentiment-analysis", model=url)

text = "text to classify"

result = classifier(text, return_all_scores=True)

label_mapping = {"LABEL_0": "Negative", "LABEL_1": "Neutral", "LABEL_2": "Positive"}
for i, predictions in enumerate(result):
  print("==================================")
  print(f"Text {i + 1}: {text}")
  for pred in predictions:
    label = label_mapping.get(pred['label'], pred['label'])
    score = pred['score']
    print(f"{label}: {score:.2%}")

Metrics and results:

It achieves the following results on the evaluation set (last epoch):

'eval_loss': 0.8515534996986389
'eval_accuracy': 0.7709153779656133
'eval_precision_per_label': [0.7665824384080859, 0.7291611185086552, 0.8197707736389684]
'eval_recall_per_label': [0.7988808426596445, 0.695630081300813, 0.8324119871981379]
'eval_f1_per_label': [0.7823984526112185, 0.7120010401768301, 0.8260430200664068]
'eval_precision_weighted': 0.7699940216435469
'eval_recall_weighted': 0.7709153779656133
'eval_f1_weighted': 0.7701923401341971
'eval_runtime': 47.0811
'eval_samples_per_second': 221.129
'eval_steps_per_second': 27.654
'epoch': 4.0

It achieves the following results on the test set:

'eval_loss': 0.8580234050750732
'eval_accuracy': 0.7714916914801652
'eval_precision_per_label': [0.7692307692307693, 0.7117024024799793, 0.8409554325662686]
'eval_recall_per_label': [0.7787552948843272, 0.7161424486612945, 0.8260371959942775]
'eval_f1_per_label': [0.7739637305699482, 0.713915522155999, 0.8334295612009238]
'eval_precision_weighted': 0.7720514465400845
'eval_recall_weighted': 0.7714916914801652
'eval_f1_weighted': 0.7717379713044402

Training Details and Procedure

Main Hyperparameters:

The following hyperparameters were used during training:

evaluation_strategy: "epoch"
learning_rate: 1e-5
per_device_train_batch_size: 8
per_device_eval_batch_size: 8
num_train_epochs: 4
optimizer: AdamW
weight_decay: 0.01
save_strategy: "epoch"
lr_scheduler_type: "linear"
warmup_steps: 820
logging_steps: 10

Preprocessing and Postprocessing:

Needed to manually map dataset creating the different sets: train 50%, validation 25%, and test 25%.
Seed=123
Num labels = 3 | srt("negative", "neutral", "positive") int(0, 1, 2)
Dynamic Padding through DataCollator was used.

Framework versions

Transformers 4.47.0
Pytorch 2.5.1+cu121
Datasets 3.2.0
Tokenizers 0.21.0

CITATION:

If you use this model, please cite the following paper:

bibitex

@inproceedings{barbieri-etal-2020-tweeteval,
    title = "{T}weet{E}val: Unified Benchmark and Comparative Evaluation for Tweet Classification",
    author = "Barbieri, Francesco  and
      Camacho-Collados, Jose  and
      Espinosa Anke, Luis  and
      Neves, Leonardo",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2020",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.findings-emnlp.148",
    doi = "10.18653/v1/2020.findings-emnlp.148",
    pages = "1644--1650"
}

More Information

Fine-tuned by Javier de la Rosa Sánchez.
javier.delarosa95@gmail.com
https://www.linkedin.com/in/delarosajav95/