Base model: roberta-large

Fine tuned for persuadee donation detection on the Persuasion For Good Dataset (Wang et al., 2019):

Given a complete dialogue from Persuasion For Good, the task is to predict the binary label:

• 0: the persuadee does not intend to donate
• 1: the persuadee intends to donate

Only persuadee utterances are input to the model for this task - persuader utterances are discarded. Each training example is the concatenation of all persuadee utterances in a single dialogue, each separated by the </s> token.

For example:

Input: <s>How are you?</s>Can you tell me more about the charity?</s>...</s>Sure, I'll donate a dollar.</s>...</s>

Label: 1

Input: <s>How are you?</s>Can you tell me more about the charity?</s>...</s>I am not interested.</s>...</s>

Label: 0

The following Dialogues were excluded:

• 146 dialogues where a donation of 0 was made at the end of the task but a non-zero amount was pledged by the persuadee in the dialogue, per the following regular expression: (?:\$(?:0\.)?[1-9]|[1-9][.0-9]*?(?: ?\$| dollars?| cents?))

Data Info:

• Training set: 587 dialogues, using actual end-task donations as labels
• Validation set: 141 dialogues, using manual donation intention labels from Persuasion For Good 'AnnSet'
• Test set: 143 dialogues, using manual donation intention labels from Persuasion For Good 'AnnSet'

Training Info:

• Loss: CrossEntropy with class weights: 1.5447 (class 0) and 0.7393 (class 1). These weights were derived from the training split.
• Early Stopping: The checkpoint with the highest validation macro f1 was selected. This occurred at step 35 (see training metrics for more detail).

Testing Info:

• Test Macro F1: 0.893
• Test Accuracy: 0.902