[Question]: TextClassifier question

[alejandrojcastaneira]

Question

Hi everyone, is there a way to perform binary classification with the TextClassifier? I have a dataset with only two classes, positive and negative, and I'm loading the corpus as a .CSV file with the following format: "text" , "label", the label can be either 0 or 1. However, the label dictionary creates two separate classes, and training begins in this manner.

After fine-tuning, the model starts predicting these two classes separately. Sometimes, a single sentence can be both, which I assume is due to the Crossentropy loss used. How can I format the training data or modify the classifier to change this behaviour? Also, would it be possible to train with BinaryCrossentropy?

[samGende]

If you want to force the Classifier to only predict one class you can init with multi_label=false:

TextClassifier(embeddings=document_embeddings, label_type=label_type, label_dictionary = label_dictionary, multi_label = False)

That should work if your .CSV only has two columns for the text and label.

You may also want to double check if the multiple labels are due to each sentence having a target label, and predicted label. When predicting the class of a sentence flair adds that prediction as a label to the Sentence object. You can specify the name of the label like this:
classifier.predict(sentence, label_name="pred")