alex2awesome / stance-detection-t5

Stance detection model distilled from a news dataset label by a larger model. The larger model was trained on a combination of stance datasets in the literature:

FNC-1 (Pomerleau and Rao, 2017), Perspectrum (Chen et al., 2019), ARC (Habernal et al., 2017), Emergent (Ferreira and Vlachos, 2016) NewsClaims (Reddy et al., 2021)7.

Achieves a .0.5712643678160919 f1-score on hand labeled indomain news data

To run:

    from transformers import AutoTokenizer, T5ForConditionalGeneration
    model = T5ForConditionalGeneration.from_pretrained('alex2awesome/stance-detection-t5')

    tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
    tokenizer_with_prefix_space = AutoTokenizer.from_pretrained(model_name_or_path, add_prefix_space=True)
    def get_tokens_as_tuple(word):
        return tuple(tokenizer_with_prefix_space([word], add_special_tokens=False).input_ids[0])

    input_ids = tokenizer(text, return_tensors="pt").input_ids
    y_pred_gen_output = model.generate(
        input_ids,
        renormalize_logits=True, 
        sequence_bias= {
            get_tokens_as_tuple('__Affirm__'):     0.143841,
            get_tokens_as_tuple('__Discuss__'):   -0.294732,
            get_tokens_as_tuple('__Neutral__'):   -0.103820,
            get_tokens_as_tuple('__Refute__'):     0.637734,
        },
    )

You can tweak the class weights yourself if you want.