Closed TianhaoFu closed 2 years ago
Thanks a lot for the question, since the MOSEI dataset is a regression dataset labeled by sentiment intensity, the label of each data instance is a float number in the range of -3.0 to 3.0 (average value from the annotators). To better evaluate the models, we can also casting the float number to the int to perform 2-class, 5-class or 7-class classification tasks. More in advance, the dataset does not only contains the sentiment intensities but also labels of the presenting of six emotions (happy, angry, etc.). More about the dataset: https://www.ml.cmu.edu/research/dap-papers/S18/dap-liang-paul-pu.pdf
Thanks a lot for the question, since the MOSEI dataset is a regression dataset labeled by sentiment intensity, the label of each data instance is a float number in the range of -3.0 to 3.0 (average value from the annotators). To better evaluate the models, we can also casting the float number to the int to perform 2-class, 5-class or 7-class classification tasks. More in advance, the dataset does not only contains the sentiment intensities but also labels of the presenting of six emotions (happy, angry, etc.). More about the dataset: https://www.ml.cmu.edu/research/dap-papers/S18/dap-liang-paul-pu.pdf
Thanks for your reply!
your mean is that mosei label maybe -2.3 or 1.3?
By the way , casting the float number to the int is the evaluation method in test data. how do you evaluate its performance in the valid data? use torch.nn.L1Loss?
Thanks:)
Yeah, it is. L1Loss will return the MAE value of the predictions, this is a traditional method to evaluate a regression dataset. As for casting the label for classification, take 2 class for example, we can use positive and negative to classify them.
For example, why some label is 1.333?
Thanks:)