Framework for Fine-tuning Transformers for Sentiment Analysis
Project description
senda
senda
is a small python package for fine-tuning transformers for
sentiment analysis (and text classification in general).
senda
builds on the excellent transformers.Trainer
API.
Installation guide
senda
can be installed from PyPI with
pip install senda
If you want the development version then install directly from GitHub.
How to use
We will show how to use senda
for fine-tuning a transformer for detecting the polarity ('positive', 'neutral' or 'negative')
of Danish Tweets. For training we use more than 5,000 Danish Tweets kindly annotated
and hosted by the Alexandra Institute (thanks!).
First, load Danish Tweets annotated with polarity.
from senda import get_danish_tweets
df_train, df_eval, df_test = get_danish_tweets()
Note, that the datasets must be DataFrames containing the columns 'text' and 'label'.
Next, instantiate the model and set up the model.
from senda import Model
m = Model(train_dataset = df_train,
eval_dataset = df_eval,
transformer = "Maltehb/danish-bert-botxo",
labels = ['negativ', 'neutral', 'positiv'],
tokenize_args = {'padding':True, 'truncation':True, 'max_length':512},
training_args = {"output_dir":'./results',
"num_train_epochs": 4,
"per_device_train_batch_size":8,
"evaluation_strategy":"steps",
"eval_steps":100,
"logging_steps":100,
"learning_rate":2e-05,
"weight_decay": 0.01,
"per_device_eval_batch_size":32,
"warmup_steps":100,
"seed":42,
"load_best_model_at_end":True,
})
Now, all there is left is to initialize a transformers.Trainer
and
train the model:
# initialize Trainer
m.init()
# run training
m.train()
The model can then be evaluated on the test set:
m.evaluate(df_test)
{'eval_loss': 0.5771588683128357, 'eval_accuracy': 0.7664399092970522, 'eval_f1': 0.7290485787279956, 'eval_runtime': 4.2016, 'eval_samples_per_second': 104.959}
Predict new observations:
text = "Sikke en dejlig dag det er i dag"
# in English: 'What a lovely day'
m.predict(text)
PredictionOutput(predictions=array([[-1.2986785 , -0.31318122, 1.2002046 ]], dtype=float32), label_ids=array([0]), metrics={'test_loss': 2.7630457878112793, 'test_accuracy': 0.0, 'test_f1': 0.0, 'test_runtime': 0.07, 'test_samples_per_second': 14.281})
m.predict(text, return_labels=True)
['positiv']
senda
model available on Huggingface
As you see, the model above achieves an accuracy of 0.77 and a macro-averaged F1-score of 0.73 on a small test data set, that Alexandra Institute provides.
The model is published on Huggingface.
Here is how to download and use the model with PyTorch:
from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline
tokenizer = AutoTokenizer.from_pretrained("pin/senda")
model = AutoModelForSequenceClassification.from_pretrained("pin/senda")
# create 'senda' sentiment analysis pipeline
senda_pipeline = pipeline('sentiment-analysis', model=model, tokenizer=tokenizer)
senda_pipeline("Sikke en dejlig dag det er i dag")
[{'label': 'positiv', 'score': 0.7678486704826355}]
The model can most certainly be improved, and we encourage all NLP-enthusiasts to try and train a better model - you can use the senda
package to do this.
senda
scope
senda
is not limited to fine-tuning transformers for Danish sentiment analysis tasks. You can use senda
for fine-tuning any transformer for any text classification task in any language.
Background
senda
is developed as a part of Ekstra Bladet’s activities on Platform Intelligence in News (PIN). PIN is an industrial research project that is carried out in collaboration between the Technical University of Denmark, University of Copenhagen and Copenhagen Business School with funding from Innovation Fund Denmark. The project runs from 2020-2023 and develops recommender systems and natural language processing systems geared for news publishing, some of which are open sourced like senda
.
Contact
We hope, that you will find senda
useful.
Please direct any questions and feedbacks to us!
If you want to contribute (which we encourage you to), open a PR.
If you encounter a bug or want to suggest an enhancement, please open an issue.
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.