roberta-large-condaqa-neg-tag-token-classification-v2

This model is a fine-tuned version of roberta-large on an unknown dataset. It achieves the following results on the evaluation set:

Loss: 0.0443
Precision: 0.0
Recall: 0.0
F1: 0.0
Accuracy: 0.9928

Model description

Negation detector. A roberta-large used for detecting negation words in sentences. A negation word will get label "Y".

Intended uses & limitations

Because the negation style in training dataset(2250 items) is not enough, maybe some kinds of negated sentences will get all "N" label.

Training and evaluation data

Using negation annotation and sentence from CondaQA and cd-sco. You can get the CondaQA dataset through both github and huggingface. As for github: https://github.com/AbhilashaRavichander/CondaQA (CondaQA) and https://github.com/mosharafhossain/negation-cue (cd-sco data). Common negation cues in CondaQA: ['halt', 'inhospitable', 'unhappy', 'unserviceable', 'dislike', 'unaware', 'unfavorable', 'barely', 'unseen', 'unoccupied', 'unreliability', 'insulator', 'stop', 'indistinguishable', 'unrestricted', 'unfairly', 'unsupervised', 'unicameral', 'forbid', 'unforgettable', 'reject', 'uneducated', 'unlimited', 'illegal', 'uncertainty', 'nonhuman', 'unborn', 'unshaven', 'uncanny', 'incomplete', 'unsure', 'unconscious', 'atypical', 'indirectly', 'unloaded', 'disadvantage', 'contrary', 'infrequent', 'unofficial', 'few', 'untouched', 'refuse', 'inequitable', 'disproportionate', 'unexpected', 'displeased', 'unpaved', 'unwieldy', 'not at all', 'absent', 'unnoticed', 'unpleasant', 'unsafe', 'unsigned', 'not', 'inaccurate', 'cannot', 'involuntary', 'unequipped', 'illiterate', 'cease', 'disagreeable', 'prohibit', 'unable', 'unstable', 'uninhabited', 'unclean', 'useless', 'disapprove', 'insensitive', 'in the absence of', 'impractical', 'unorthodox', 'untreated', 'unsuccessful', 'unwitting', 'unfashionable', 'disagreement', 'unmyelinated', 'unfortunate', 'unknown', 'ineffective', 'a lack of', 'instead of', 'refused', 'illegitimate', 'little', 'unpaid', 'fail', 'unintentionally', 'unglazed', "didn't", 'unprocessed', 'inability', 'undeveloped', 'exclude', 'neither', 'except', 'unequivocal', 'unconventional', 'incorrectly', 'unconditional', 'prevent', 'dissimilar', 'uncommon', 'inorganic', 'unquestionable', 'uncoated', 'unassisted', 'unprecedented', 'nonviolent', 'unarmed', 'unpopular', 'inadequate', 'uncomfortable', 'unwilling', 'unaffected', 'unfaithful', 'nobody', 'loss', 'without', 'undamaged', 'nothing', 'could not', 'impossible to', 'unaccompanied', 'unlike', 'oppose', 'compromising', 'unmarried', 'rarely', 'unlighted', 'inexperienced', 'rather than', 'unrelated', 'untied', 'dishonest', 'insecure', 'uneven', 'harmless', 'avoid', 'with the exception of', 'no', 'undefeated', 'no longer', 'inadvertently', 'absence', 'lack', 'unconnected', 'unfinished', 'invalid', 'unnecessary', 'invisibility', 'unusual', 'none', 'incredulous', 'impossible', 'never', 'untrained', 'incorrect', 'immobility', 'unclear', 'impartial', 'unlucky', 'deny', 'uncertain', 'hardly', 'unsaturated', 'informal', 'irregular', 'dissatisfaction'] More information needed

Training procedure

Use code from huggingface source(token-classification).

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 2e-05
train_batch_size: 16
eval_batch_size: 8
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 10.0

Training results

Framework versions

Transformers 4.25.0.dev0
Pytorch 1.10.1
Datasets 2.6.1
Tokenizers 0.13.1