diwank/silicone-deberta-pair
deberta-base-based dialog acts classifier. Trained on the balanced variant of the silicone-merged dataset: a simplified merged dialog act data from datasets in the silicone collection.
Takes two sentences as inputs (one previous and one current utterance of a dialog). The previous sentence can be an empty string if this is the first utterance of a speaker in a dialog. Outputs one of 11 labels:
(0, 'acknowledge')
(1, 'answer')
(2, 'backchannel')
(3, 'reply_yes')
(4, 'exclaim')
(5, 'say')
(6, 'reply_no')
(7, 'hold')
(8, 'ask')
(9, 'intent')
(10, 'ask_yes_no')
Example:
from simpletransformers.classification import (
ClassificationModel, ClassificationArgs
)
model = ClassificationModel("deberta", "diwank/silicone-deberta-pair")
convert_to_label = lambda n: [
['acknowledge',
'answer',
'backchannel',
'reply_yes',
'exclaim',
'say',
'reply_no',
'hold',
'ask',
'intent',
'ask_yes_no'
][i] for i in n
]
predictions, raw_outputs = model.predict([["Say what is the meaning of life?", "I dont know"]])
convert_to_label(predictions) # answer